audit-tools 0.55.15 → 0.55.17
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/dist/audit/cli/auditStep.d.ts.map +1 -1
- package/dist/audit/cli/auditStep.js +4 -3
- package/dist/audit/cli/auditStep.js.map +1 -1
- package/dist/audit/cli/charterComparisonPrompt.d.ts.map +1 -1
- package/dist/audit/cli/charterComparisonPrompt.js +8 -1
- package/dist/audit/cli/charterComparisonPrompt.js.map +1 -1
- package/dist/audit/cli/conceptualDispatch.d.ts.map +1 -1
- package/dist/audit/cli/conceptualDispatch.js +9 -5
- package/dist/audit/cli/conceptualDispatch.js.map +1 -1
- package/dist/audit/cli/confirmIntentStep.d.ts.map +1 -1
- package/dist/audit/cli/confirmIntentStep.js +0 -1
- package/dist/audit/cli/confirmIntentStep.js.map +1 -1
- package/dist/audit/cli/dispatch/hostHandoff.d.ts +86 -1
- package/dist/audit/cli/dispatch/hostHandoff.d.ts.map +1 -1
- package/dist/audit/cli/dispatch/hostHandoff.js +460 -161
- package/dist/audit/cli/dispatch/hostHandoff.js.map +1 -1
- package/dist/audit/cli/fanoutLanes.d.ts +1 -1
- package/dist/audit/cli/fanoutLanes.d.ts.map +1 -1
- package/dist/audit/cli/fanoutLanes.js +4 -2
- package/dist/audit/cli/fanoutLanes.js.map +1 -1
- package/dist/audit/cli/nextStepCommand.d.ts.map +1 -1
- package/dist/audit/cli/nextStepCommand.js +17 -15
- package/dist/audit/cli/nextStepCommand.js.map +1 -1
- package/dist/audit/cli/nextStepHelpers.d.ts.map +1 -1
- package/dist/audit/cli/nextStepHelpers.js +4 -3
- package/dist/audit/cli/nextStepHelpers.js.map +1 -1
- package/dist/audit/cli/prompts.d.ts +2 -0
- package/dist/audit/cli/prompts.d.ts.map +1 -1
- package/dist/audit/cli/prompts.js +12 -1
- package/dist/audit/cli/prompts.js.map +1 -1
- package/dist/audit/cli/semanticReviewStep.d.ts.map +1 -1
- package/dist/audit/cli/semanticReviewStep.js +71 -6
- package/dist/audit/cli/semanticReviewStep.js.map +1 -1
- package/dist/audit/cli/submitCommand.d.ts +26 -0
- package/dist/audit/cli/submitCommand.d.ts.map +1 -0
- package/dist/audit/cli/submitCommand.js +79 -0
- package/dist/audit/cli/submitCommand.js.map +1 -0
- package/dist/audit/cli/validateResultsCommand.d.ts +11 -0
- package/dist/audit/cli/validateResultsCommand.d.ts.map +1 -1
- package/dist/audit/cli/validateResultsCommand.js +56 -4
- package/dist/audit/cli/validateResultsCommand.js.map +1 -1
- package/dist/audit/cli.d.ts.map +1 -1
- package/dist/audit/cli.js +4 -1
- package/dist/audit/cli.js.map +1 -1
- package/dist/audit/contracts/findingContractPrompt.d.ts +229 -0
- package/dist/audit/contracts/findingContractPrompt.d.ts.map +1 -1
- package/dist/audit/contracts/findingContractPrompt.js +31 -20
- package/dist/audit/contracts/findingContractPrompt.js.map +1 -1
- package/dist/audit/decompose/buildStructureDecomposition.d.ts +11 -5
- package/dist/audit/decompose/buildStructureDecomposition.d.ts.map +1 -1
- package/dist/audit/decompose/buildStructureDecomposition.js +17 -8
- package/dist/audit/decompose/buildStructureDecomposition.js.map +1 -1
- package/dist/audit/extractors/designAssessment.d.ts.map +1 -1
- package/dist/audit/extractors/designAssessment.js.map +1 -1
- package/dist/audit/extractors/graph.d.ts +1 -1
- package/dist/audit/extractors/graph.js +1 -1
- package/dist/audit/extractors/graphSignals.d.ts +8 -0
- package/dist/audit/extractors/graphSignals.d.ts.map +1 -1
- package/dist/audit/extractors/graphSignals.js +12 -1
- package/dist/audit/extractors/graphSignals.js.map +1 -1
- package/dist/audit/extractors/pathPatterns.d.ts +3 -0
- package/dist/audit/extractors/pathPatterns.d.ts.map +1 -1
- package/dist/audit/extractors/pathPatterns.js +9 -4
- package/dist/audit/extractors/pathPatterns.js.map +1 -1
- package/dist/audit/extractors/surfaces.d.ts.map +1 -1
- package/dist/audit/extractors/surfaces.js +6 -1
- package/dist/audit/extractors/surfaces.js.map +1 -1
- package/dist/audit/io/runArtifacts.d.ts +6 -0
- package/dist/audit/io/runArtifacts.d.ts.map +1 -1
- package/dist/audit/io/runArtifacts.js +9 -0
- package/dist/audit/io/runArtifacts.js.map +1 -1
- package/dist/audit/orchestrator/advance.d.ts.map +1 -1
- package/dist/audit/orchestrator/advance.js +3 -0
- package/dist/audit/orchestrator/advance.js.map +1 -1
- package/dist/audit/orchestrator/architectureDiscovery.d.ts +9 -0
- package/dist/audit/orchestrator/architectureDiscovery.d.ts.map +1 -1
- package/dist/audit/orchestrator/architectureDiscovery.js +11 -1
- package/dist/audit/orchestrator/architectureDiscovery.js.map +1 -1
- package/dist/audit/orchestrator/charterPackets.d.ts +36 -7
- package/dist/audit/orchestrator/charterPackets.d.ts.map +1 -1
- package/dist/audit/orchestrator/charterPackets.js +88 -21
- package/dist/audit/orchestrator/charterPackets.js.map +1 -1
- package/dist/audit/orchestrator/dependencyMap.d.ts.map +1 -1
- package/dist/audit/orchestrator/dependencyMap.js +16 -13
- package/dist/audit/orchestrator/dependencyMap.js.map +1 -1
- package/dist/audit/orchestrator/dependencySlices.d.ts.map +1 -1
- package/dist/audit/orchestrator/dependencySlices.js +26 -11
- package/dist/audit/orchestrator/dependencySlices.js.map +1 -1
- package/dist/audit/orchestrator/designReviewPrompt.d.ts.map +1 -1
- package/dist/audit/orchestrator/designReviewPrompt.js +38 -6
- package/dist/audit/orchestrator/designReviewPrompt.js.map +1 -1
- package/dist/audit/orchestrator/intentCheckpointExecutor.d.ts.map +1 -1
- package/dist/audit/orchestrator/intentCheckpointExecutor.js +3 -2
- package/dist/audit/orchestrator/intentCheckpointExecutor.js.map +1 -1
- package/dist/audit/orchestrator/state.js +1 -1
- package/dist/audit/orchestrator/state.js.map +1 -1
- package/dist/audit/orchestrator/systemicChallengeExecutor.d.ts +9 -0
- package/dist/audit/orchestrator/systemicChallengeExecutor.d.ts.map +1 -1
- package/dist/audit/orchestrator/systemicChallengeExecutor.js +12 -2
- package/dist/audit/orchestrator/systemicChallengeExecutor.js.map +1 -1
- package/dist/audit/orchestrator/unitBuilder.d.ts.map +1 -1
- package/dist/audit/orchestrator/unitBuilder.js +29 -8
- package/dist/audit/orchestrator/unitBuilder.js.map +1 -1
- package/dist/audit/reporting/criticalFlowFallbackPrompt.d.ts +6 -0
- package/dist/audit/reporting/criticalFlowFallbackPrompt.d.ts.map +1 -1
- package/dist/audit/reporting/criticalFlowFallbackPrompt.js +55 -34
- package/dist/audit/reporting/criticalFlowFallbackPrompt.js.map +1 -1
- package/dist/audit/systemic/reviewFileMap.d.ts +92 -45
- package/dist/audit/systemic/reviewFileMap.d.ts.map +1 -1
- package/dist/audit/systemic/reviewFileMap.js +171 -94
- package/dist/audit/systemic/reviewFileMap.js.map +1 -1
- package/dist/audit/systemic/secondOrderAdversaryPrompt.d.ts +15 -6
- package/dist/audit/systemic/secondOrderAdversaryPrompt.d.ts.map +1 -1
- package/dist/audit/systemic/secondOrderAdversaryPrompt.js +119 -36
- package/dist/audit/systemic/secondOrderAdversaryPrompt.js.map +1 -1
- package/dist/audit/types/designAssessment.d.ts +11 -0
- package/dist/audit/types/designAssessment.d.ts.map +1 -1
- package/dist/audit/types/designAssessment.js +23 -1
- package/dist/audit/types/designAssessment.js.map +1 -1
- package/dist/remediate/index.d.ts.map +1 -1
- package/dist/remediate/index.js +44 -0
- package/dist/remediate/index.js.map +1 -1
- package/dist/remediate/steps/dispatch/hostHandoff.d.ts +29 -0
- package/dist/remediate/steps/dispatch/hostHandoff.d.ts.map +1 -1
- package/dist/remediate/steps/dispatch/hostHandoff.js +63 -4
- package/dist/remediate/steps/dispatch/hostHandoff.js.map +1 -1
- package/dist/remediate/steps/submitResult.d.ts +13 -0
- package/dist/remediate/steps/submitResult.d.ts.map +1 -0
- package/dist/remediate/steps/submitResult.js +24 -0
- package/dist/remediate/steps/submitResult.js.map +1 -0
- package/dist/shared/decompose/charterExtraction.d.ts +117 -17
- package/dist/shared/decompose/charterExtraction.d.ts.map +1 -1
- package/dist/shared/decompose/charterExtraction.js +27 -3
- package/dist/shared/decompose/charterExtraction.js.map +1 -1
- package/dist/shared/decompose/systemicChallenge.d.ts.map +1 -1
- package/dist/shared/decompose/systemicChallenge.js +4 -1
- package/dist/shared/decompose/systemicChallenge.js.map +1 -1
- package/dist/shared/index.d.ts +2 -2
- package/dist/shared/index.d.ts.map +1 -1
- package/dist/shared/index.js +2 -2
- package/dist/shared/index.js.map +1 -1
- package/dist/shared/mcp/submitServer.d.ts +26 -0
- package/dist/shared/mcp/submitServer.d.ts.map +1 -0
- package/dist/shared/mcp/submitServer.js +74 -0
- package/dist/shared/mcp/submitServer.js.map +1 -0
- package/dist/shared/submission/hostHandoffCore.d.ts +10 -0
- package/dist/shared/submission/hostHandoffCore.d.ts.map +1 -1
- package/dist/shared/submission/hostHandoffCore.js +19 -0
- package/dist/shared/submission/hostHandoffCore.js.map +1 -1
- package/dist/shared/submission/submissionClassifier.d.ts +7 -0
- package/dist/shared/submission/submissionClassifier.d.ts.map +1 -1
- package/dist/shared/submission/submissionClassifier.js +9 -0
- package/dist/shared/submission/submissionClassifier.js.map +1 -1
- package/dist/shared/submission/submissionScan.d.ts +7 -0
- package/dist/shared/submission/submissionScan.d.ts.map +1 -1
- package/dist/shared/submission/submissionScan.js +6 -2
- package/dist/shared/submission/submissionScan.js.map +1 -1
- package/package.json +2 -1
- package/scripts/audit/postinstall.mjs +21 -7
- package/scripts/remediate/postinstall.mjs +44 -1
- package/scripts/shared/install-host-assets.mjs +134 -1
- package/wrapper/audit-code-wrapper-lib.mjs +9 -46
|
@@ -3,9 +3,9 @@ import { ActiveReviewRunSchema } from "../../contracts/wrapperResponse.js";
|
|
|
3
3
|
import { sourceTreeInstruction } from "../../io/runSnapshot.js";
|
|
4
4
|
import { CURRENT_TASK_FILENAME } from "../../supervisor/operatorHandoff.js";
|
|
5
5
|
import { mkdir } from "node:fs/promises";
|
|
6
|
-
import { join } from "node:path";
|
|
6
|
+
import { join, relative, resolve } from "node:path";
|
|
7
7
|
import { lensDefinition } from "../../types.js";
|
|
8
|
-
import { createMemoizedSourceReader, enumerateTrackedFilePaths, deriveLaneDemand, appendSubmissionEvent, assertSubmissionRunId, SUBMISSION_LEDGER_EVENT_CONTRACT_VERSION, bindingIdentity, compareCodeUnits, contentSha256, firstDuplicateIdentity, hasExactKeys, hostHandoffResultPath, identityFailureDiagnostic, isFileMissingError, isJsonParseError, isRecord, isSha256, LaneDemandSchema, parseAllWorkloadItems, parseWorkloadEnvelope, bindWorkerPrompt, deriveResultId, workerPromptBindingHolds, enrichMissingSubmissionIssues, readTrailingSubmissionRefusals, readJsonFile, repoRelativePath, requireNonEmptyString, resolveContainedPath, resolveHostHandoffPaths, resultMapIdentity, scanBoundSubmission, siblingLockPath, sameStrings, stableStringify, groundFinding, withFileLock, writeBlockedStepContract, writeJsonFile, } from "audit-tools/shared";
|
|
8
|
+
import { createMemoizedSourceReader, enumerateTrackedFilePaths, deriveLaneDemand, appendSubmissionEvent, assertSubmissionRunId, SUBMISSION_LEDGER_EVENT_CONTRACT_VERSION, bindingIdentity, compareCodeUnits, contentSha256, firstDuplicateIdentity, hasExactKeys, hostHandoffResultPath, workerPromptPath, identityFailureDiagnostic, isFileMissingError, isJsonParseError, isRecord, isSha256, LaneDemandSchema, parseAllWorkloadItems, parseWorkloadEnvelope, bindWorkerPrompt, deriveResultId, workerPromptBindingHolds, enrichMissingSubmissionIssues, readTrailingSubmissionRefusals, readJsonFile, repoRelativePath, requireNonEmptyString, resolveContainedPath, resolveHostHandoffPaths, resultMapIdentity, scanBoundSubmission, siblingLockPath, sameStrings, stableStringify, groundFinding, withFileLock, writeBlockedStepContract, writeTextFile, writeJsonFile, toPosixPath, } from "audit-tools/shared";
|
|
9
9
|
import { findingContractPromptLines } from "../../contracts/findingContractPrompt.js";
|
|
10
10
|
import { WORKER_REFUSED_FINDING_VERDICTS, WORKER_REFUSED_LOCATION_FIELDS, WorkerFindingSchema, } from "../../contracts/workerSchemas.js";
|
|
11
11
|
import { AuditResultSchema, AuditTaskSchema, } from "../../types.js";
|
|
@@ -14,7 +14,7 @@ import { validateOneAuditResult, formatAuditResultIssues,
|
|
|
14
14
|
// `validateVerification` build their allowed set through the same function, so
|
|
15
15
|
// they cannot disagree about which `file_paths` entries are inside the packet.
|
|
16
16
|
verificationAllowedPaths, normalizeCoveragePath, } from "../../validation/auditResults.js";
|
|
17
|
-
import { reviewWaveClosedPath } from "../../io/runArtifacts.js";
|
|
17
|
+
import { hostCoherenceGroupsPath, reviewWaveClosedPath } from "../../io/runArtifacts.js";
|
|
18
18
|
// The lane token itself is single-sourced where the steward lane is minted:
|
|
19
19
|
// `LENS_VERIFICATION_TAG` (`orchestrator/selectiveDeepening/shared.ts`). Every
|
|
20
20
|
// reader compiles against that one declaration, so a rename moves them together
|
|
@@ -22,6 +22,8 @@ import { reviewWaveClosedPath } from "../../io/runArtifacts.js";
|
|
|
22
22
|
// is what a hand-copied literal here produced, twice (this boundary and the
|
|
23
23
|
// audit-results validator).
|
|
24
24
|
import { LENS_VERIFICATION_TAG, } from "../../orchestrator/selectiveDeepening/shared.js";
|
|
25
|
+
import { buildTaskAffinityGraph } from "../../orchestrator/taskAffinityGraph.js";
|
|
26
|
+
import { buildTaskCoherencePartition } from "../../orchestrator/partitionTaskGraph.js";
|
|
25
27
|
// v1alpha2 (the emitted-lane demand ranking): each work item's `metadata`
|
|
26
28
|
// carries the shared `demand` ranking (size / complexity / risk) beside its
|
|
27
29
|
// token estimate, in place of the retired `{complexity, risk}` pair whose
|
|
@@ -44,7 +46,17 @@ import { LENS_VERIFICATION_TAG, } from "../../orchestrator/selectiveDeepening/sh
|
|
|
44
46
|
// the result template, its identity values filled in by the tool, and the
|
|
45
47
|
// prompt digest covers the text ABOVE that template. A v1alpha3 document refuses
|
|
46
48
|
// closed as stale — the same `next-step` re-prepares it.
|
|
47
|
-
|
|
49
|
+
//
|
|
50
|
+
// v1alpha5 (the 0.55.15 self-audit): the prompt and its result template name
|
|
51
|
+
// the result path ABSOLUTE, so the persisted template a work item is checked
|
|
52
|
+
// against changed shape. A v1alpha4 document refuses closed as stale — the same
|
|
53
|
+
// `next-step` re-prepares it.
|
|
54
|
+
//
|
|
55
|
+
// v1alpha6 (decision 5 of the 2026-11-06 dogfood audit): each fact stated once.
|
|
56
|
+
// The template no longer names the path (line 2 of the body does), and its
|
|
57
|
+
// wording depends on the lane, so the persisted template changed shape again. A
|
|
58
|
+
// v1alpha5 document refuses closed as stale — the same `next-step` re-prepares it.
|
|
59
|
+
const WORKLOAD_CONTRACT_VERSION = "audit-host-workload/v1alpha6";
|
|
48
60
|
const RESULT_MAP_CONTRACT_VERSION = "audit-host-result-map/v1alpha1";
|
|
49
61
|
const RESULT_CONTRACT_VERSION = "audit-host-result/v1alpha1";
|
|
50
62
|
/**
|
|
@@ -362,11 +374,13 @@ function isVerificationLane(tags) {
|
|
|
362
374
|
return tags?.includes(LENS_VERIFICATION_TAG) ?? false;
|
|
363
375
|
}
|
|
364
376
|
/**
|
|
365
|
-
* The
|
|
366
|
-
*
|
|
367
|
-
*
|
|
377
|
+
* The tail every work-item prompt ends with. The destination is NOT repeated
|
|
378
|
+
* here: line 2 of the body states it, once (decision 5, cut C1). The base lane's
|
|
379
|
+
* key set is stated by the template itself ("Keep exactly these keys", cut C2);
|
|
380
|
+
* the steward lane adds `verification`, so its body keeps the result-contract
|
|
381
|
+
* line and its tail does not claim an exact key set.
|
|
368
382
|
*/
|
|
369
|
-
function renderResultTemplate(runId, workItemId,
|
|
383
|
+
function renderResultTemplate(runId, workItemId, promptDigest, verificationLane) {
|
|
370
384
|
const template = {
|
|
371
385
|
contract_version: RESULT_CONTRACT_VERSION,
|
|
372
386
|
result_id: deriveResultId(workItemId, promptDigest),
|
|
@@ -377,7 +391,9 @@ function renderResultTemplate(runId, workItemId, resultPath, promptDigest) {
|
|
|
377
391
|
findings: [],
|
|
378
392
|
};
|
|
379
393
|
return [
|
|
380
|
-
|
|
394
|
+
verificationLane
|
|
395
|
+
? "When you finish, submit this JSON as line 2 says. Keep the five identity values as they are. Fill in `file_coverage` and `findings`:"
|
|
396
|
+
: "When you finish, submit this JSON as line 2 says. Keep exactly these keys, and keep the five identity values as they are. Fill in `file_coverage` and `findings`:",
|
|
381
397
|
"```json",
|
|
382
398
|
JSON.stringify(template, null, 2),
|
|
383
399
|
"```",
|
|
@@ -402,23 +418,13 @@ function lensGuidanceLines(lens) {
|
|
|
402
418
|
`Do not report under this lens: ${definition.do_not_report}`,
|
|
403
419
|
];
|
|
404
420
|
}
|
|
405
|
-
function buildPrompt(task,
|
|
421
|
+
function buildPrompt(task, destination, workloadPath, sourceRoot) {
|
|
406
422
|
// A `"selective"` lane's file list is NOT inlined. Its assignment is the whole
|
|
407
423
|
// surface its lens was applied to, which runs to hundreds of files on a real
|
|
408
424
|
// run, so inlining it would make the prompt mostly a path list and the prompt
|
|
409
425
|
// DIGEST a function of the surface's size. The list and its per-file metrics
|
|
410
426
|
// ride the workload the tool already writes, and the prompt names where.
|
|
411
427
|
const selective = task.coverage_policy === "selective";
|
|
412
|
-
const assignment = stableStringify({
|
|
413
|
-
...(selective
|
|
414
|
-
? { surface_file_count: task.file_paths.length }
|
|
415
|
-
: { file_line_counts: task.file_line_counts, files: task.file_paths }),
|
|
416
|
-
lens: task.lens,
|
|
417
|
-
pass_id: task.pass_id,
|
|
418
|
-
rationale: task.rationale,
|
|
419
|
-
task_id: task.task_id,
|
|
420
|
-
unit_id: task.unit_id,
|
|
421
|
-
});
|
|
422
428
|
// LANE-AWARE, because the two lanes' contracts genuinely differ: the steward
|
|
423
429
|
// lane is INSTRUCTED (by its task rationale) to return verification metadata,
|
|
424
430
|
// and the base lane is not. Rendering one envelope sentence for both would
|
|
@@ -427,16 +433,35 @@ function buildPrompt(task, resultPath, workloadPath, sourceRoot) {
|
|
|
427
433
|
// demands. The ask and the envelope are the same fact, so they are stated
|
|
428
434
|
// from the same predicate.
|
|
429
435
|
const verificationLane = isVerificationLane(task.tags);
|
|
436
|
+
// Each fact once (decision 5): `file_line_counts` already names every file
|
|
437
|
+
// (cut C3), and the base lane's reader uses neither `task_id` (the template's
|
|
438
|
+
// `work_item_id`) nor `pass_id` (`pass:<lens>`) (cut C4). The steward keeps
|
|
439
|
+
// both: each of its `verification.followup_tasks` entries must carry them.
|
|
440
|
+
const assignment = stableStringify({
|
|
441
|
+
...(selective
|
|
442
|
+
? { surface_file_count: task.file_paths.length }
|
|
443
|
+
: { file_line_counts: task.file_line_counts }),
|
|
444
|
+
lens: task.lens,
|
|
445
|
+
...(verificationLane ? { pass_id: task.pass_id, task_id: task.task_id } : {}),
|
|
446
|
+
rationale: task.rationale,
|
|
447
|
+
unit_id: task.unit_id,
|
|
448
|
+
});
|
|
430
449
|
return [
|
|
431
450
|
"Perform the bounded semantic audit work item below.",
|
|
432
451
|
// THE DESTINATION FIRST, on its own line. It used to be one field inside a
|
|
433
452
|
// single long JSON blob, which buries the one fact a reader must not have to
|
|
434
453
|
// search for: a result written anywhere else is never ingested.
|
|
435
|
-
|
|
454
|
+
// ABSOLUTE (`workerPromptPath`): the worker's working directory is the
|
|
455
|
+
// host's choice, so no path here may depend on it.
|
|
456
|
+
destination,
|
|
436
457
|
// WHERE TO READ, stated rather than left to the host's working directory:
|
|
437
458
|
// the run audits a frozen snapshot, and the live tree may have moved since
|
|
438
459
|
// the run started. Line counts and cited text are checked against this tree.
|
|
439
460
|
sourceTreeInstruction(sourceRoot),
|
|
461
|
+
// WHAT ELSE TO WRITE: nothing. Workers in the self-audit wrote scratch
|
|
462
|
+
// files to the OS temp directory and stalled on the permission prompt each
|
|
463
|
+
// one raised; the work needs no file but the result.
|
|
464
|
+
"Write no other file: no scratch, temporary or helper file anywhere. Both paths above are absolute and do not depend on your working directory.",
|
|
440
465
|
// LANE-AWARE, because "review every listed file" is true of the base lane
|
|
441
466
|
// and false of a steward under selective coverage — whose whole task is to
|
|
442
467
|
// decide what is worth opening.
|
|
@@ -453,23 +478,41 @@ function buildPrompt(task, resultPath, workloadPath, sourceRoot) {
|
|
|
453
478
|
assignment,
|
|
454
479
|
"```",
|
|
455
480
|
...lensGuidanceLines(task.lens),
|
|
456
|
-
|
|
457
|
-
|
|
458
|
-
|
|
459
|
-
|
|
460
|
-
|
|
461
|
-
|
|
462
|
-
|
|
463
|
-
|
|
464
|
-
|
|
481
|
+
// `verification` is REQUIRED on the steward lane, and the prompt states ONE
|
|
482
|
+
// rule. It used to say "exactly … and verification" and then call the field
|
|
483
|
+
// optional, which is a contradiction a reader cannot obey both halves of.
|
|
484
|
+
// Required is the honest reading: a steward returns `findings: []`, so a
|
|
485
|
+
// steward with no `verification` carries no answer at all and is
|
|
486
|
+
// indistinguishable from a lane that failed. The base lane has no such
|
|
487
|
+
// line: its template states the exact key set (decision 5, cut C2).
|
|
488
|
+
...(verificationLane
|
|
489
|
+
? ["Result contract: audit-host-result/v1alpha1 with exactly result_id, run_id, work_item_id, prompt_sha256, file_coverage, findings, and verification in addition to contract_version. On this lane verification is REQUIRED."]
|
|
490
|
+
: []),
|
|
465
491
|
"Each file_coverage entry must contain exactly path, reviewed_lines, and total_lines.",
|
|
466
492
|
// The finding contract is CARRIED, not referenced: it is rendered from the
|
|
467
493
|
// very schema ingestion enforces, so a host never has to remember or fetch it.
|
|
468
494
|
...findingContractPromptLines(),
|
|
469
|
-
"
|
|
495
|
+
"Never supply a `grounding` field: the tool computes it at ingest from your quoted_text.",
|
|
470
496
|
...(verificationLane ? verificationContractPromptLines() : []),
|
|
497
|
+
// Stated ONCE, last, so it covers every rule above it (decision 5, cut C7).
|
|
498
|
+
"A result that breaks any rule above is rejected whole.",
|
|
471
499
|
].join("\n");
|
|
472
500
|
}
|
|
501
|
+
/**
|
|
502
|
+
* Line 2 of every work-item prompt: how the result reaches the tool, and the one
|
|
503
|
+
* statement of the absolute destination (owner decision, 2026-11-07). The
|
|
504
|
+
* submit tool comes first because it checks the result before it writes it, and
|
|
505
|
+
* it needs the root and the work-item id, so the line states both. The file path
|
|
506
|
+
* stays as the fallback for a host without the tool. A shell is excluded by
|
|
507
|
+
* name: the dogfood audit lost a backslash in a quoted span that way.
|
|
508
|
+
*/
|
|
509
|
+
function destinationLine(root, workItemId, resultPath) {
|
|
510
|
+
return (`Submit one JSON object with the submit_result tool (root: ${workerPromptPath(root, ".")}, work_item_id: ${workItemId}). ` +
|
|
511
|
+
`If you do not have that tool, write the object with a file-write tool to this exact absolute path: ${workerPromptPath(root, resultPath)}. ` +
|
|
512
|
+
"Do not write it through a shell.");
|
|
513
|
+
}
|
|
514
|
+
/** The first line of the steward's verification contract; only a steward body holds it. */
|
|
515
|
+
const VERIFICATION_CONTRACT_LEAD = "This work item is a LENS STEWARD VERIFICATION task, so it must also carry a `verification` object.";
|
|
473
516
|
/**
|
|
474
517
|
* The `verification` contract rendered into a steward-lane work item's prompt.
|
|
475
518
|
*
|
|
@@ -482,7 +525,7 @@ function buildPrompt(task, resultPath, workloadPath, sourceRoot) {
|
|
|
482
525
|
*/
|
|
483
526
|
function verificationContractPromptLines() {
|
|
484
527
|
return [
|
|
485
|
-
|
|
528
|
+
VERIFICATION_CONTRACT_LEAD,
|
|
486
529
|
`verification must contain exactly ${VERIFICATION_CONTRACT_KEYS.join(", ")} — all ${String(VERIFICATION_CONTRACT_KEYS.length)}, with no extra key: verified and needs_followup are booleans, concerns, coverage_concerns and confidence_concerns are arrays of non-empty strings (a genuine "nothing to report" is an empty array, not an omitted field), selection_rationale is a non-empty string, and followup_tasks is an array of objects.`,
|
|
487
530
|
"verification.selection_rationale states how you chose which surface files to open and which to leave: name the signals you followed and say what you decided was not worth opening. A file you did not open is not a coverage failure, but an unexplained choice is.",
|
|
488
531
|
`Each verification.followup_tasks entry must contain exactly ${VERIFICATION_FOLLOWUP_KEYS.join(", ")} — all ${String(VERIFICATION_FOLLOWUP_KEYS.length)}, with no extra key: file_paths is a non-empty array of non-empty repo-relative strings, each naming a file on THIS work item's assigned surface (scope.files) — a file you did not open is still on the surface and may be named; lens must be the lens of THIS task; task_id, unit_id, pass_id and rationale must be non-empty strings.`,
|
|
@@ -636,7 +679,7 @@ function buildWorkItem(paths, task) {
|
|
|
636
679
|
const resultPath = resultPathFor(paths, task.task_id);
|
|
637
680
|
// The SAME path the caller writes the workload to, so a selective lane is told
|
|
638
681
|
// where its surface actually is rather than where it is expected to be.
|
|
639
|
-
const prompt = bindWorkerPrompt(buildPrompt(task, resultPath, paths.workloadPath, boundarySourceRoot(paths)), (digest) => renderResultTemplate(paths.runId, task.task_id,
|
|
682
|
+
const prompt = bindWorkerPrompt(buildPrompt(task, destinationLine(paths.root, task.task_id, resultPath), paths.workloadPath, boundarySourceRoot(paths)), (digest) => renderResultTemplate(paths.runId, task.task_id, digest, isVerificationLane(task.tags)));
|
|
640
683
|
return {
|
|
641
684
|
id: task.task_id,
|
|
642
685
|
lens: task.lens,
|
|
@@ -805,11 +848,20 @@ export async function prepareAuditHostHandoff(params) {
|
|
|
805
848
|
await writeAcceptedResults(paths, accepted);
|
|
806
849
|
await writeJsonFile(paths.workloadPath, workload);
|
|
807
850
|
await writeJsonFile(paths.resultMapPath, resultMap);
|
|
851
|
+
const coherenceGroups = {
|
|
852
|
+
contract_version: COHERENCE_GROUPS_CONTRACT_VERSION,
|
|
853
|
+
run_id: params.runId,
|
|
854
|
+
groups: coherenceGroupsFor(allWorkItems, params.coherenceGroups),
|
|
855
|
+
};
|
|
856
|
+
const coherenceGroupsPath = hostCoherenceGroupsPath(params.artifactsDir, params.runId);
|
|
857
|
+
await writeJsonFile(coherenceGroupsPath, coherenceGroups);
|
|
808
858
|
return {
|
|
809
859
|
workload,
|
|
810
860
|
result_map: resultMap,
|
|
811
861
|
workload_path: paths.workloadPath,
|
|
812
862
|
result_map_path: paths.resultMapPath,
|
|
863
|
+
coherence_groups: coherenceGroups,
|
|
864
|
+
coherence_groups_path: coherenceGroupsPath,
|
|
813
865
|
};
|
|
814
866
|
});
|
|
815
867
|
}
|
|
@@ -825,7 +877,7 @@ export async function prepareAuditHostHandoff(params) {
|
|
|
825
877
|
function isSurfaceFileMetrics(value) {
|
|
826
878
|
return AuditTaskSchema.shape.file_metrics.safeParse(value).success;
|
|
827
879
|
}
|
|
828
|
-
function parseWorkItem(value, runId) {
|
|
880
|
+
function parseWorkItem(value, runId, root) {
|
|
829
881
|
// `file_metrics` is the ONE optional scope key: a selective (steward) lane
|
|
830
882
|
// carries it, every complete lane omits it. It is added to the expected set
|
|
831
883
|
// only when the document supplies it, exactly as the result envelope treats
|
|
@@ -854,7 +906,18 @@ function parseWorkItem(value, runId) {
|
|
|
854
906
|
!hasExactKeys(value.prompt, ["sha256", "text"]) ||
|
|
855
907
|
!isSha256(value.prompt.sha256) ||
|
|
856
908
|
typeof value.prompt.text !== "string" ||
|
|
857
|
-
!workerPromptBindingHolds({ text: value.prompt.text, sha256: value.prompt.sha256 },
|
|
909
|
+
!workerPromptBindingHolds({ text: value.prompt.text, sha256: value.prompt.sha256 },
|
|
910
|
+
// The lane is read from the digest-bound BODY: only a steward body states
|
|
911
|
+
// the verification contract, and the body cannot change without the digest.
|
|
912
|
+
(digest) => renderResultTemplate(runId, value.id, digest,
|
|
913
|
+
// A WHOLE line: a rationale is free text inside the assignment JSON, so a
|
|
914
|
+
// substring match would let it flip the lane.
|
|
915
|
+
value.prompt.text.split("\n").includes(VERIFICATION_CONTRACT_LEAD))) ||
|
|
916
|
+
// Line 2 states the absolute destination under the root the workload was
|
|
917
|
+
// prepared at. Re-derived under the CURRENT root, a root that really moved
|
|
918
|
+
// no longer matches, and the workload is refused as stale (see
|
|
919
|
+
// `workerPromptPath`) rather than sending readers to the old tree.
|
|
920
|
+
value.prompt.text.split("\n")[1] !== destinationLine(root, value.id, value.result_path) ||
|
|
858
921
|
!isRecord(value.scope) ||
|
|
859
922
|
!hasExactKeys(value.scope, [
|
|
860
923
|
"files",
|
|
@@ -942,7 +1005,7 @@ class StaleAuditHostTaskBindingsError extends Error {
|
|
|
942
1005
|
this.name = "StaleAuditHostTaskBindingsError";
|
|
943
1006
|
}
|
|
944
1007
|
}
|
|
945
|
-
function parseWorkload(value, runId) {
|
|
1008
|
+
function parseWorkload(value, runId, root) {
|
|
946
1009
|
// Envelope + all-items parsing is the CORE's scaffolding; the audit draw
|
|
947
1010
|
// selects only its own contract version and item parser.
|
|
948
1011
|
const envelope = parseWorkloadEnvelope(value, {
|
|
@@ -962,7 +1025,7 @@ function parseWorkload(value, runId) {
|
|
|
962
1025
|
}
|
|
963
1026
|
throw bindingFailure("workload_binding", "Invalid audit host workload");
|
|
964
1027
|
}
|
|
965
|
-
const workItems = parseAllWorkloadItems(envelope.rawItems, (raw) => parseWorkItem(raw, runId));
|
|
1028
|
+
const workItems = parseAllWorkloadItems(envelope.rawItems, (raw) => parseWorkItem(raw, runId, root));
|
|
966
1029
|
if (workItems === null) {
|
|
967
1030
|
throw bindingFailure("workload_binding", "Invalid audit host work item");
|
|
968
1031
|
}
|
|
@@ -1086,33 +1149,43 @@ function validateHandoffBinding(paths, workload, resultMap, taskBindings) {
|
|
|
1086
1149
|
}
|
|
1087
1150
|
return items;
|
|
1088
1151
|
}
|
|
1089
|
-
/**
|
|
1090
|
-
|
|
1152
|
+
/**
|
|
1153
|
+
* The run the current-review-run pointer names, or `undefined` when there is
|
|
1154
|
+
* no usable pointer. The pointer schema accepts strings; only the existing
|
|
1155
|
+
* run-id grammar may authorize a directory segment, so an unusable pointer is
|
|
1156
|
+
* not publication.
|
|
1157
|
+
*/
|
|
1158
|
+
async function readActiveRunId(artifactsDir) {
|
|
1091
1159
|
let run;
|
|
1092
1160
|
try {
|
|
1093
|
-
run = ActiveReviewRunSchema.safeParse(await readJsonFile(join(
|
|
1161
|
+
run = ActiveReviewRunSchema.safeParse(await readJsonFile(join(artifactsDir, "dispatch", CURRENT_TASK_FILENAME)));
|
|
1094
1162
|
}
|
|
1095
1163
|
catch (error) {
|
|
1096
1164
|
if (isFileMissingError(error) || isJsonParseError(error))
|
|
1097
|
-
return
|
|
1165
|
+
return undefined;
|
|
1098
1166
|
throw error;
|
|
1099
1167
|
}
|
|
1100
1168
|
if (!run.success)
|
|
1101
|
-
return
|
|
1102
|
-
// The pointer schema accepts strings; only the existing run-id grammar may
|
|
1103
|
-
// authorize a directory segment. An unusable pointer is not publication.
|
|
1169
|
+
return undefined;
|
|
1104
1170
|
try {
|
|
1105
1171
|
assertSubmissionRunId(run.data.run_id);
|
|
1106
1172
|
}
|
|
1107
1173
|
catch {
|
|
1108
|
-
return
|
|
1174
|
+
return undefined;
|
|
1109
1175
|
}
|
|
1110
|
-
|
|
1176
|
+
return run.data.run_id;
|
|
1177
|
+
}
|
|
1178
|
+
/** Snapshot current publication under its existing lock; release before planning. */
|
|
1179
|
+
export async function readPublishedAuditTaskIds(params) {
|
|
1180
|
+
const runId = await readActiveRunId(params.artifactsDir);
|
|
1181
|
+
if (runId === undefined)
|
|
1182
|
+
return new Set();
|
|
1183
|
+
const paths = resolveBoundaryPaths({ root: params.root, sourceRoot: params.sourceRoot, artifactsDir: params.artifactsDir, runId });
|
|
1111
1184
|
return withAcceptedResultsLock(paths, params.logger, async () => {
|
|
1112
1185
|
let items;
|
|
1113
1186
|
let bindings;
|
|
1114
1187
|
try {
|
|
1115
|
-
const workload = parseWorkload(await readJsonFile(paths.workloadPath), paths.runId);
|
|
1188
|
+
const workload = parseWorkload(await readJsonFile(paths.workloadPath), paths.runId, paths.root);
|
|
1116
1189
|
const resultMap = parseResultMap(await readJsonFile(paths.resultMapPath), paths.runId);
|
|
1117
1190
|
bindings = parseTaskBindings(await readJsonFile(paths.taskBindingsPath), paths.runId);
|
|
1118
1191
|
items = validateHandoffBinding(paths, workload, resultMap, bindings);
|
|
@@ -1462,6 +1535,325 @@ function evictInvalidatedEntries(loaded, params) {
|
|
|
1462
1535
|
});
|
|
1463
1536
|
return { kept: { ...loaded, entries: kept }, issues };
|
|
1464
1537
|
}
|
|
1538
|
+
/**
|
|
1539
|
+
* The ONE judgment of one bound submission: read it at its bound path, parse
|
|
1540
|
+
* the envelope against its binding, convert it, ground its cited text, and put
|
|
1541
|
+
* it through the per-result validator. Ingestion ACCEPTS on `ok`; the
|
|
1542
|
+
* read-only check ({@link checkAuditHostResult}) only reports — so a host that
|
|
1543
|
+
* checks a result before ingestion is told exactly what ingestion would say.
|
|
1544
|
+
*/
|
|
1545
|
+
async function judgeAuditHostSubmission(params) {
|
|
1546
|
+
const { paths, entry, item, binding } = params;
|
|
1547
|
+
const outcome = await scanBoundSubmission({
|
|
1548
|
+
root: paths.root,
|
|
1549
|
+
artifactsDir: paths.artifactsDir,
|
|
1550
|
+
workItemId: entry.work_item_id,
|
|
1551
|
+
resultPath: entry.result_path,
|
|
1552
|
+
parse: (value) => {
|
|
1553
|
+
const parsed = parseHostResult(value, paths.runId, item, binding);
|
|
1554
|
+
return parsed.ok ? { ok: true, parsed: parsed.result } : parsed;
|
|
1555
|
+
},
|
|
1556
|
+
resultId: (result) => result.result_id,
|
|
1557
|
+
seen: params.seen,
|
|
1558
|
+
messages: auditScanMessages(entry.work_item_id),
|
|
1559
|
+
source: params.source,
|
|
1560
|
+
});
|
|
1561
|
+
if (!outcome.ok)
|
|
1562
|
+
return { ok: false, issue: outcome.issue };
|
|
1563
|
+
const result = outcome.parsed;
|
|
1564
|
+
const converted = toAuditResult(result, binding);
|
|
1565
|
+
if (!converted.ok) {
|
|
1566
|
+
return {
|
|
1567
|
+
ok: false,
|
|
1568
|
+
issue: {
|
|
1569
|
+
code: "submission_contract_invalid",
|
|
1570
|
+
check: "result_schema",
|
|
1571
|
+
message: `work item '${entry.work_item_id}' submitted a result that does not convert to ` +
|
|
1572
|
+
`an AuditResult: ${converted.detail}`,
|
|
1573
|
+
work_item_id: entry.work_item_id,
|
|
1574
|
+
result_path: entry.result_path,
|
|
1575
|
+
},
|
|
1576
|
+
};
|
|
1577
|
+
}
|
|
1578
|
+
// S7 quote-and-verify: the tool re-reads each cited span from disk, stamps
|
|
1579
|
+
// the verdict, and writes each entry's line numbers from where its quote
|
|
1580
|
+
// occurs. It NEVER rejects — a quote that is absent or occurs more than
|
|
1581
|
+
// once rides through as `ungrounded` and synthesis surfaces it under
|
|
1582
|
+
// "Ungrounded Findings (not confirmed)"; refusing here would discard the
|
|
1583
|
+
// whole submission over one bad citation. It mutates the CONVERTED copy
|
|
1584
|
+
// only: `result` stays the host's submission, which `result_sha256` hashes.
|
|
1585
|
+
// It runs BEFORE validation, so the derived span meets the same coverage
|
|
1586
|
+
// rule here that the downstream batch gate applies: a span that rule would
|
|
1587
|
+
// refuse is refused while the item can still be resubmitted, never
|
|
1588
|
+
// accepted and then replayed into a failing gate on every later step.
|
|
1589
|
+
for (const finding of converted.auditResult.findings) {
|
|
1590
|
+
await groundFinding(boundarySourceRoot(paths), finding, params.readSource, await params.knownPaths());
|
|
1591
|
+
}
|
|
1592
|
+
// VALIDATE BEFORE ACCEPT. The conversion above proves only the envelope
|
|
1593
|
+
// contract (`FindingSchema` admits an evidence-less finding); these are the
|
|
1594
|
+
// per-result rules the downstream batch gate applies, applied HERE so an
|
|
1595
|
+
// error-severity issue never reaches the accepted pair. A rejected item is
|
|
1596
|
+
// simply never in the ledger, so the corrected file at the same bound path
|
|
1597
|
+
// is re-read on the next fold — acceptance used to be terminal instead, and
|
|
1598
|
+
// a failed batch gate then wedged the run permanently.
|
|
1599
|
+
//
|
|
1600
|
+
// Orphans (task pruned by a re-plan) pass through UNVALIDATED with the same
|
|
1601
|
+
// stderr notice the batch gate uses — never newly rejected: refusing one
|
|
1602
|
+
// here would strand it outside the append-only ledger entirely.
|
|
1603
|
+
if (!params.activeTaskIds.has(entry.work_item_id)) {
|
|
1604
|
+
process.stderr.write(`audit host-handoff ingest: result for '${entry.work_item_id}' is not in the ` +
|
|
1605
|
+
`active task manifest (orphaned by re-planning); retained in the accepted pair ` +
|
|
1606
|
+
`but skipped at the validation gate\n`);
|
|
1607
|
+
return { ok: true, result, auditResult: converted.auditResult, warnings: [] };
|
|
1608
|
+
}
|
|
1609
|
+
const validationIssues = validateOneAuditResult(converted.auditResult, [...params.auditTasks], {
|
|
1610
|
+
lineIndex: params.lineIndex,
|
|
1611
|
+
});
|
|
1612
|
+
const errors = validationIssues.filter((issue) => issue.severity === "error");
|
|
1613
|
+
if (errors.length > 0) {
|
|
1614
|
+
return {
|
|
1615
|
+
ok: false,
|
|
1616
|
+
issue: {
|
|
1617
|
+
code: "result_validation_failed",
|
|
1618
|
+
check: "result_validation",
|
|
1619
|
+
message: `work item '${entry.work_item_id}' failed audit-results validation ` +
|
|
1620
|
+
`(${errors.length} error(s)); fix the result file at its bound path and call next-step again: ` +
|
|
1621
|
+
formatAuditResultIssues(errors),
|
|
1622
|
+
work_item_id: entry.work_item_id,
|
|
1623
|
+
result_path: entry.result_path,
|
|
1624
|
+
},
|
|
1625
|
+
};
|
|
1626
|
+
}
|
|
1627
|
+
// Warnings are NOT rejections: an accepted result never refused anything,
|
|
1628
|
+
// so a warning must never reach the rejection-classified issue list — the
|
|
1629
|
+
// ONE ledger recorder would otherwise record kind:'rejected' for a result
|
|
1630
|
+
// that was accepted, manufacturing a repair story that never happened.
|
|
1631
|
+
// They ride the separate advisory channel instead (rendered for the
|
|
1632
|
+
// operator; never counted as a submission that could not be accepted).
|
|
1633
|
+
const warnings = validationIssues
|
|
1634
|
+
.filter((issue) => issue.severity === "warning")
|
|
1635
|
+
.map((warning) => ({
|
|
1636
|
+
work_item_id: entry.work_item_id,
|
|
1637
|
+
result_path: entry.result_path,
|
|
1638
|
+
message: `${warning.message} (${warning.field})`,
|
|
1639
|
+
}));
|
|
1640
|
+
return { ok: true, result, auditResult: converted.auditResult, warnings };
|
|
1641
|
+
}
|
|
1642
|
+
/**
|
|
1643
|
+
* The run's published binding set — workload, result map, task bindings — read
|
|
1644
|
+
* and cross-checked, or the CLASSIFIED refusal ingestion reports for it. One
|
|
1645
|
+
* reader for both doors, so the read-only check refuses a stale or corrupt set
|
|
1646
|
+
* exactly as ingestion does, never as a stack. Infrastructure failures stay
|
|
1647
|
+
* strict.
|
|
1648
|
+
*/
|
|
1649
|
+
async function readHandoffBindingSet(paths) {
|
|
1650
|
+
try {
|
|
1651
|
+
const workload = parseWorkload(await readJsonFile(paths.workloadPath), paths.runId, paths.root);
|
|
1652
|
+
const resultMap = parseResultMap(await readJsonFile(paths.resultMapPath), paths.runId);
|
|
1653
|
+
const taskBindings = parseTaskBindings(await readJsonFile(paths.taskBindingsPath), paths.runId);
|
|
1654
|
+
const items = validateHandoffBinding(paths, workload, resultMap, taskBindings);
|
|
1655
|
+
return { ok: true, resultMap, taskBindings, items };
|
|
1656
|
+
}
|
|
1657
|
+
catch (error) {
|
|
1658
|
+
if (error instanceof StaleAuditHostWorkloadError ||
|
|
1659
|
+
error instanceof StaleAuditHostTaskBindingsError ||
|
|
1660
|
+
error instanceof AuditHostBindingError)
|
|
1661
|
+
return { ok: false, refusal: error };
|
|
1662
|
+
if (isJsonParseError(error)) {
|
|
1663
|
+
return { ok: false, refusal: new AuditHostBindingError("workload_binding", error.message) };
|
|
1664
|
+
}
|
|
1665
|
+
throw error;
|
|
1666
|
+
}
|
|
1667
|
+
}
|
|
1668
|
+
/**
|
|
1669
|
+
* Judge one host result file by the rules ingestion applies, WITHOUT accepting
|
|
1670
|
+
* it: nothing is written, no lock is held, the ledger is only read. The run is
|
|
1671
|
+
* named by where the file lives (`<artifacts>/runs/<run id>/host-results/…`),
|
|
1672
|
+
* and the file must be the bound result path of one of that run's work items.
|
|
1673
|
+
*/
|
|
1674
|
+
export async function checkAuditHostResult(params) {
|
|
1675
|
+
const absolute = resolve(params.resultPath);
|
|
1676
|
+
const segments = toPosixPath(relative(resolve(params.artifactsDir, "runs"), absolute)).split("/");
|
|
1677
|
+
if (segments.length !== 3 || segments[1] !== "host-results" || segments[0] === "..") {
|
|
1678
|
+
throw new Error(`${toPosixPath(absolute)} is not a bound host result path. A semantic-review result is written at the ` +
|
|
1679
|
+
"`result_path` its work item binds in the published host-workload.json, under " +
|
|
1680
|
+
`${toPosixPath(join(resolve(params.artifactsDir), "runs", "<run id>", "host-results"))}/; ` +
|
|
1681
|
+
"write it there and validate that file.");
|
|
1682
|
+
}
|
|
1683
|
+
const runId = segments[0];
|
|
1684
|
+
const paths = resolveBoundaryPaths({
|
|
1685
|
+
root: params.root,
|
|
1686
|
+
sourceRoot: params.sourceRoot,
|
|
1687
|
+
artifactsDir: params.artifactsDir,
|
|
1688
|
+
runId,
|
|
1689
|
+
});
|
|
1690
|
+
let bindingSet;
|
|
1691
|
+
try {
|
|
1692
|
+
bindingSet = await readHandoffBindingSet(paths);
|
|
1693
|
+
}
|
|
1694
|
+
catch (error) {
|
|
1695
|
+
if (!isFileMissingError(error))
|
|
1696
|
+
throw error;
|
|
1697
|
+
throw new Error(`run ${runId} has no published workload (${toPosixPath(paths.workloadPath)} is missing); ` +
|
|
1698
|
+
"run next-step to publish one, then write each result at the result_path it binds.");
|
|
1699
|
+
}
|
|
1700
|
+
if (!bindingSet.ok) {
|
|
1701
|
+
// The SAME classified refusal ingestion reports for this binding set.
|
|
1702
|
+
return {
|
|
1703
|
+
run_id: runId,
|
|
1704
|
+
already_accepted: false,
|
|
1705
|
+
issues: [{ code: bindingSet.refusal.code, check: bindingSet.refusal.check, message: bindingSet.refusal.message }],
|
|
1706
|
+
validation_warnings: [],
|
|
1707
|
+
};
|
|
1708
|
+
}
|
|
1709
|
+
const entry = bindingSet.resultMap.entries.find((candidate) => resolve(paths.root, candidate.result_path) === absolute);
|
|
1710
|
+
if (entry === undefined) {
|
|
1711
|
+
throw new Error(`${toPosixPath(absolute)} is not the bound result path of any work item in run ${runId}. ` +
|
|
1712
|
+
`Each work item in ${toPosixPath(paths.workloadPath)} names its own result_path; write the result there and validate that file.`);
|
|
1713
|
+
}
|
|
1714
|
+
const judged = await judgeBoundEntry({ ...params, paths, bindingSet, entry });
|
|
1715
|
+
return checkReport(runId, entry, judged);
|
|
1716
|
+
}
|
|
1717
|
+
/**
|
|
1718
|
+
* The judgment both doors share once a run's binding set and one of its entries
|
|
1719
|
+
* are resolved: the ledger read, the duplicate predicate, and the ONE
|
|
1720
|
+
* submission judgment. `source` carries `submit`'s in-memory document.
|
|
1721
|
+
*/
|
|
1722
|
+
async function judgeBoundEntry(params) {
|
|
1723
|
+
const { paths, bindingSet, entry } = params;
|
|
1724
|
+
const item = bindingSet.items.get(entry.work_item_id);
|
|
1725
|
+
const binding = bindingSet.taskBindings.get(entry.work_item_id);
|
|
1726
|
+
if (item === undefined || binding === undefined) {
|
|
1727
|
+
throw new Error(`work item '${entry.work_item_id}' has a result path in run ${paths.runId} but no work item or task binding; ` +
|
|
1728
|
+
"run next-step to publish a current workload.");
|
|
1729
|
+
}
|
|
1730
|
+
const accepted = await loadAcceptedResults(paths.acceptedLedgerPath, paths.runId);
|
|
1731
|
+
const alreadyAccepted = accepted.entries.some((candidate) => bindingIdentity(candidate) === bindingIdentity(entry));
|
|
1732
|
+
// An accepted binding's own id is not a duplicate of itself: the check judges
|
|
1733
|
+
// the bytes, and says separately that ingestion will not re-read them.
|
|
1734
|
+
const otherIds = new Set(accepted.entries
|
|
1735
|
+
.filter((candidate) => candidate.work_item_id !== entry.work_item_id)
|
|
1736
|
+
.map((candidate) => candidate.result_id));
|
|
1737
|
+
let knownPaths;
|
|
1738
|
+
const judged = await judgeAuditHostSubmission({
|
|
1739
|
+
paths,
|
|
1740
|
+
entry,
|
|
1741
|
+
item,
|
|
1742
|
+
binding,
|
|
1743
|
+
seen: (resultId) => otherIds.has(resultId),
|
|
1744
|
+
activeTaskIds: new Set(params.auditTasks.map((task) => task.task_id)),
|
|
1745
|
+
auditTasks: params.auditTasks,
|
|
1746
|
+
lineIndex: params.lineIndex,
|
|
1747
|
+
readSource: createMemoizedSourceReader(),
|
|
1748
|
+
knownPaths: async () => (knownPaths ??= await enumerateTrackedFilePaths(boundarySourceRoot(paths))),
|
|
1749
|
+
source: params.source,
|
|
1750
|
+
});
|
|
1751
|
+
return { alreadyAccepted, judged };
|
|
1752
|
+
}
|
|
1753
|
+
function checkReport(runId, entry, { alreadyAccepted, judged }) {
|
|
1754
|
+
return {
|
|
1755
|
+
run_id: runId,
|
|
1756
|
+
work_item_id: entry.work_item_id,
|
|
1757
|
+
result_path: entry.result_path,
|
|
1758
|
+
already_accepted: alreadyAccepted,
|
|
1759
|
+
issues: judged.ok ? [] : [judged.issue],
|
|
1760
|
+
validation_warnings: judged.ok ? judged.warnings : [],
|
|
1761
|
+
};
|
|
1762
|
+
}
|
|
1763
|
+
/**
|
|
1764
|
+
* Judge a submitted result for one work item of the CURRENT run and, only when
|
|
1765
|
+
* ingestion would accept it, write it atomically to that item's bound result
|
|
1766
|
+
* path (decision 4, 2026-11-06). The reader never names a path: the tool
|
|
1767
|
+
* derives it from the work-item id. A refused result writes nothing, and the
|
|
1768
|
+
* report carries the same refusal ingestion would give, so the reader fixes
|
|
1769
|
+
* the document and submits again. No lock is held: the write is the one a
|
|
1770
|
+
* reader made before, and the next-step fold still ingests the file under the
|
|
1771
|
+
* artifact-tree lock. Bindings are read on every call, never cached, so an item
|
|
1772
|
+
* a re-plan retired fails with the classified refusal.
|
|
1773
|
+
*/
|
|
1774
|
+
export async function submitAuditHostResult(params) {
|
|
1775
|
+
const runId = await readActiveRunId(params.artifactsDir);
|
|
1776
|
+
if (runId === undefined) {
|
|
1777
|
+
throw new Error("no audit run has published a workload (the current-review-run pointer is missing or unusable); " +
|
|
1778
|
+
"run next-step to publish one, then submit each result by its work-item id.");
|
|
1779
|
+
}
|
|
1780
|
+
const paths = resolveBoundaryPaths({
|
|
1781
|
+
root: params.root,
|
|
1782
|
+
sourceRoot: params.sourceRoot,
|
|
1783
|
+
artifactsDir: params.artifactsDir,
|
|
1784
|
+
runId,
|
|
1785
|
+
});
|
|
1786
|
+
const bindingSet = await readHandoffBindingSet(paths);
|
|
1787
|
+
if (!bindingSet.ok) {
|
|
1788
|
+
return {
|
|
1789
|
+
run_id: runId,
|
|
1790
|
+
already_accepted: false,
|
|
1791
|
+
issues: [{ code: bindingSet.refusal.code, check: bindingSet.refusal.check, message: bindingSet.refusal.message }],
|
|
1792
|
+
validation_warnings: [],
|
|
1793
|
+
};
|
|
1794
|
+
}
|
|
1795
|
+
const entry = bindingSet.resultMap.entries.find((candidate) => candidate.work_item_id === params.workItemId);
|
|
1796
|
+
if (entry === undefined) {
|
|
1797
|
+
throw new Error(`'${params.workItemId}' is not a work item of the current run ${runId}. ` +
|
|
1798
|
+
`Submit with the work-item id your prompt names; the run's items are listed in ${toPosixPath(paths.workloadPath)}.`);
|
|
1799
|
+
}
|
|
1800
|
+
const judgment = await judgeBoundEntry({ ...params, paths, bindingSet, entry });
|
|
1801
|
+
if (judgment.alreadyAccepted) {
|
|
1802
|
+
throw new Error(`work item '${entry.work_item_id}' is already accepted in run ${runId}, and ingestion does not read an accepted result again. ` +
|
|
1803
|
+
`To replace it, run \`audit-code unaccept-results --work-item ${entry.work_item_id}\` first, then submit again.`);
|
|
1804
|
+
}
|
|
1805
|
+
if (judgment.judged.ok) {
|
|
1806
|
+
await writeTextFile(resolveContainedPath(paths.root, entry.result_path, `result path for ${entry.work_item_id}`), params.source);
|
|
1807
|
+
}
|
|
1808
|
+
return checkReport(runId, entry, judgment);
|
|
1809
|
+
}
|
|
1810
|
+
/** The set-level scheduling facts beside a published workload. */
|
|
1811
|
+
const COHERENCE_GROUPS_CONTRACT_VERSION = "audit-host-coherence-groups/v1alpha1";
|
|
1812
|
+
/**
|
|
1813
|
+
* The groups to publish: the caller's, which must partition EXACTLY the
|
|
1814
|
+
* published items (a group naming an unpublished item, or an item in no group,
|
|
1815
|
+
* would describe a different publish), or one singleton per item.
|
|
1816
|
+
*/
|
|
1817
|
+
function coherenceGroupsFor(items, supplied) {
|
|
1818
|
+
if (supplied === undefined) {
|
|
1819
|
+
return items.map((item) => ({
|
|
1820
|
+
id: item.id,
|
|
1821
|
+
work_item_ids: [item.id],
|
|
1822
|
+
token_estimate: item.metadata.token_estimate,
|
|
1823
|
+
}));
|
|
1824
|
+
}
|
|
1825
|
+
const members = supplied.flatMap((group) => group.work_item_ids).sort(compareCodeUnits);
|
|
1826
|
+
const published = items.map((item) => item.id);
|
|
1827
|
+
if (stableStringify(members) !== stableStringify(published)) {
|
|
1828
|
+
throw new Error("audit host coherence groups do not partition the published work items exactly");
|
|
1829
|
+
}
|
|
1830
|
+
return supplied;
|
|
1831
|
+
}
|
|
1832
|
+
/**
|
|
1833
|
+
* Content-coherence groups over the work items a wave PUBLISHES: the shared
|
|
1834
|
+
* task-draw partition (`buildTaskCoherencePartition`) of the affinity graph
|
|
1835
|
+
* built from exactly those tasks. Items in one group share files, a unit, a
|
|
1836
|
+
* flow, a directory or graph adjacency, so one lane taking them together reads
|
|
1837
|
+
* each file once.
|
|
1838
|
+
*
|
|
1839
|
+
* SET-LEVEL, deliberately, and never a field on a work item: a group is a
|
|
1840
|
+
* property of the whole published set, so on an item it would rewrite that
|
|
1841
|
+
* item whenever a peer is added or accepted — and an unchanged published item
|
|
1842
|
+
* must re-publish byte-identical. Metadata only: order, grouping and
|
|
1843
|
+
* parallelism stay the host's choice.
|
|
1844
|
+
*/
|
|
1845
|
+
export function buildAuditHostCoherenceGroups(tasks, graphBundle) {
|
|
1846
|
+
const partition = tasks.length === 0
|
|
1847
|
+
? { packets: [] }
|
|
1848
|
+
: buildTaskCoherencePartition(buildTaskAffinityGraph(tasks, graphBundle === undefined ? {} : { graphBundle }));
|
|
1849
|
+
const groups = partition.packets
|
|
1850
|
+
.map((packet) => {
|
|
1851
|
+
const ids = [...packet.task_ids].sort(compareCodeUnits);
|
|
1852
|
+
return { id: ids[0], work_item_ids: ids, token_estimate: packet.token_estimate };
|
|
1853
|
+
})
|
|
1854
|
+
.sort((left, right) => compareCodeUnits(left.id, right.id));
|
|
1855
|
+
return groups;
|
|
1856
|
+
}
|
|
1465
1857
|
export async function ingestAuditHostResults(params) {
|
|
1466
1858
|
const paths = resolveBoundaryPaths(params);
|
|
1467
1859
|
return withAcceptedResultsLock(paths, params.logger, async (loaded) => {
|
|
@@ -1497,27 +1889,10 @@ export async function ingestAuditHostResults(params) {
|
|
|
1497
1889
|
raw_issues: [...evictionIssues(), error],
|
|
1498
1890
|
validation_warnings: [],
|
|
1499
1891
|
});
|
|
1500
|
-
|
|
1501
|
-
|
|
1502
|
-
|
|
1503
|
-
|
|
1504
|
-
try {
|
|
1505
|
-
workload = parseWorkload(await readJsonFile(paths.workloadPath), params.runId);
|
|
1506
|
-
resultMap = parseResultMap(await readJsonFile(paths.resultMapPath), params.runId);
|
|
1507
|
-
taskBindings = parseTaskBindings(await readJsonFile(paths.taskBindingsPath), params.runId);
|
|
1508
|
-
items = validateHandoffBinding(paths, workload, resultMap, taskBindings);
|
|
1509
|
-
}
|
|
1510
|
-
catch (error) {
|
|
1511
|
-
if (error instanceof StaleAuditHostWorkloadError ||
|
|
1512
|
-
error instanceof StaleAuditHostTaskBindingsError ||
|
|
1513
|
-
error instanceof AuditHostBindingError)
|
|
1514
|
-
return stale(error);
|
|
1515
|
-
if (isJsonParseError(error)) {
|
|
1516
|
-
return stale(new AuditHostBindingError("workload_binding", error.message));
|
|
1517
|
-
}
|
|
1518
|
-
// Infrastructure failures and corrupt accepted ledgers remain strict.
|
|
1519
|
-
throw error;
|
|
1520
|
-
}
|
|
1892
|
+
const bindingSet = await readHandoffBindingSet(paths);
|
|
1893
|
+
if (!bindingSet.ok)
|
|
1894
|
+
return stale(bindingSet.refusal);
|
|
1895
|
+
const { resultMap, taskBindings, items } = bindingSet;
|
|
1521
1896
|
const acceptedBindings = new Set(accepted.entries.map(bindingIdentity));
|
|
1522
1897
|
const resultIds = new Set(accepted.entries.map((entry) => entry.result_id));
|
|
1523
1898
|
const additions = [];
|
|
@@ -1535,102 +1910,26 @@ export async function ingestAuditHostResults(params) {
|
|
|
1535
1910
|
const binding = taskBindings.get(entry.work_item_id);
|
|
1536
1911
|
if (item === undefined || binding === undefined)
|
|
1537
1912
|
continue;
|
|
1538
|
-
const
|
|
1539
|
-
|
|
1540
|
-
|
|
1541
|
-
|
|
1542
|
-
|
|
1543
|
-
|
|
1544
|
-
|
|
1545
|
-
return parsed.ok ? { ok: true, parsed: parsed.result } : parsed;
|
|
1546
|
-
},
|
|
1547
|
-
resultId: (result) => result.result_id,
|
|
1548
|
-
// CHECK only. The id is consumed further down, after conversion,
|
|
1549
|
-
// validation and grounding — a result refused there must stay re-submittable.
|
|
1913
|
+
const judged = await judgeAuditHostSubmission({
|
|
1914
|
+
paths,
|
|
1915
|
+
entry,
|
|
1916
|
+
item,
|
|
1917
|
+
binding,
|
|
1918
|
+
// CHECK only. The id is consumed below, after the judgment — a result
|
|
1919
|
+
// refused there must stay re-submittable.
|
|
1550
1920
|
seen: (resultId) => resultIds.has(resultId),
|
|
1551
|
-
|
|
1921
|
+
activeTaskIds,
|
|
1922
|
+
auditTasks: params.auditTasks,
|
|
1923
|
+
lineIndex: params.lineIndex,
|
|
1924
|
+
readSource,
|
|
1925
|
+
knownPaths: async () => (knownPaths ??= await enumerateTrackedFilePaths(boundarySourceRoot(paths))),
|
|
1552
1926
|
});
|
|
1553
|
-
if (!
|
|
1554
|
-
issues.push(
|
|
1555
|
-
continue;
|
|
1556
|
-
}
|
|
1557
|
-
const result = outcome.parsed;
|
|
1558
|
-
const converted = toAuditResult(result, binding);
|
|
1559
|
-
if (!converted.ok) {
|
|
1560
|
-
issues.push({
|
|
1561
|
-
code: "submission_contract_invalid",
|
|
1562
|
-
check: "result_schema",
|
|
1563
|
-
message: `work item '${entry.work_item_id}' submitted a result that does not convert to ` +
|
|
1564
|
-
`an AuditResult: ${converted.detail}`,
|
|
1565
|
-
work_item_id: entry.work_item_id,
|
|
1566
|
-
result_path: entry.result_path,
|
|
1567
|
-
});
|
|
1927
|
+
if (!judged.ok) {
|
|
1928
|
+
issues.push(judged.issue);
|
|
1568
1929
|
continue;
|
|
1569
1930
|
}
|
|
1570
|
-
|
|
1571
|
-
|
|
1572
|
-
// occurs. It NEVER rejects — a quote that is absent or occurs more than
|
|
1573
|
-
// once rides through as `ungrounded` and synthesis surfaces it under
|
|
1574
|
-
// "Ungrounded Findings (not confirmed)"; refusing here would discard the
|
|
1575
|
-
// whole submission over one bad citation. It mutates the CONVERTED copy
|
|
1576
|
-
// only: `result` stays the host's submission, which `result_sha256` hashes.
|
|
1577
|
-
// It runs BEFORE validation, so the derived span meets the same coverage
|
|
1578
|
-
// rule here that the downstream batch gate applies: a span that rule would
|
|
1579
|
-
// refuse is refused while the item can still be resubmitted, never
|
|
1580
|
-
// accepted and then replayed into a failing gate on every later step.
|
|
1581
|
-
for (const finding of converted.auditResult.findings) {
|
|
1582
|
-
knownPaths ??= await enumerateTrackedFilePaths(boundarySourceRoot(paths));
|
|
1583
|
-
await groundFinding(boundarySourceRoot(paths), finding, readSource, knownPaths);
|
|
1584
|
-
}
|
|
1585
|
-
// VALIDATE BEFORE ACCEPT. The conversion above proves only the envelope
|
|
1586
|
-
// contract (`FindingSchema` admits an evidence-less finding); these are the
|
|
1587
|
-
// per-result rules the downstream batch gate applies, applied HERE so an
|
|
1588
|
-
// error-severity issue never reaches the accepted pair. A rejected item is
|
|
1589
|
-
// simply never in the ledger, so the corrected file at the same bound path
|
|
1590
|
-
// is re-read on the next fold — acceptance used to be terminal instead, and
|
|
1591
|
-
// a failed batch gate then wedged the run permanently.
|
|
1592
|
-
//
|
|
1593
|
-
// Orphans (task pruned by a re-plan) pass through UNVALIDATED with the same
|
|
1594
|
-
// stderr notice the batch gate uses — never newly rejected: refusing one
|
|
1595
|
-
// here would strand it outside the append-only ledger entirely.
|
|
1596
|
-
if (!activeTaskIds.has(entry.work_item_id)) {
|
|
1597
|
-
process.stderr.write(`audit host-handoff ingest: result for '${entry.work_item_id}' is not in the ` +
|
|
1598
|
-
`active task manifest (orphaned by re-planning); retained in the accepted pair ` +
|
|
1599
|
-
`but skipped at the validation gate\n`);
|
|
1600
|
-
}
|
|
1601
|
-
else {
|
|
1602
|
-
const validationIssues = validateOneAuditResult(converted.auditResult, [
|
|
1603
|
-
...params.auditTasks,
|
|
1604
|
-
], {
|
|
1605
|
-
lineIndex: params.lineIndex,
|
|
1606
|
-
});
|
|
1607
|
-
const errors = validationIssues.filter((issue) => issue.severity === "error");
|
|
1608
|
-
if (errors.length > 0) {
|
|
1609
|
-
issues.push({
|
|
1610
|
-
code: "result_validation_failed",
|
|
1611
|
-
check: "result_validation",
|
|
1612
|
-
message: `work item '${entry.work_item_id}' failed audit-results validation ` +
|
|
1613
|
-
`(${errors.length} error(s)); fix the result file at its bound path and call next-step again: ` +
|
|
1614
|
-
formatAuditResultIssues(errors),
|
|
1615
|
-
work_item_id: entry.work_item_id,
|
|
1616
|
-
result_path: entry.result_path,
|
|
1617
|
-
});
|
|
1618
|
-
continue;
|
|
1619
|
-
}
|
|
1620
|
-
// Warnings are NOT rejections: an accepted result never refused anything,
|
|
1621
|
-
// so a warning must never reach the rejection-classified issue list — the
|
|
1622
|
-
// ONE ledger recorder would otherwise record kind:'rejected' for a result
|
|
1623
|
-
// that was accepted, manufacturing a repair story that never happened.
|
|
1624
|
-
// They ride the separate advisory channel instead (rendered for the
|
|
1625
|
-
// operator; never counted as a submission that could not be accepted).
|
|
1626
|
-
validation_warnings.push(...validationIssues
|
|
1627
|
-
.filter((issue) => issue.severity === "warning")
|
|
1628
|
-
.map((warning) => ({
|
|
1629
|
-
work_item_id: entry.work_item_id,
|
|
1630
|
-
result_path: entry.result_path,
|
|
1631
|
-
message: `${warning.message} (${warning.field})`,
|
|
1632
|
-
})));
|
|
1633
|
-
}
|
|
1931
|
+
validation_warnings.push(...judged.warnings);
|
|
1932
|
+
const { result, auditResult } = judged;
|
|
1634
1933
|
resultIds.add(result.result_id);
|
|
1635
1934
|
// construction-site: AuditResult
|
|
1636
1935
|
// construction-site: Finding (the `findings` array; `lens` defaults from the enclosing contract, already spread in)
|
|
@@ -1641,7 +1940,7 @@ export async function ingestAuditHostResults(params) {
|
|
|
1641
1940
|
result_id: result.result_id,
|
|
1642
1941
|
result_sha256: contentSha256(result),
|
|
1643
1942
|
result,
|
|
1644
|
-
audit_result:
|
|
1943
|
+
audit_result: auditResult,
|
|
1645
1944
|
});
|
|
1646
1945
|
}
|
|
1647
1946
|
// Binding reads, validation and this accepted pair write share ONE hold.
|