crysk 0.3.2 → 0.4.0
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 +52 -32
- package/dist/SKILL-12as0f3q.md +74 -0
- package/dist/SKILL-2h445snn.md +140 -0
- package/dist/SKILL-6vy283pt.md +140 -0
- package/dist/SKILL-dsnffnfp.md +140 -0
- package/dist/SKILL-ebn0fqaq.md +140 -0
- package/dist/SKILL-gnrbph72.md +140 -0
- package/dist/SKILL-kq0xejn3.md +553 -0
- package/dist/SKILL-qaqsbygw.md +66 -0
- package/dist/SKILL-vws0aqb8.md +553 -0
- package/dist/ci-admission-a6d0wk1c.md +128 -0
- package/dist/ci-admission-t0f5qwth.md +128 -0
- package/dist/cli.js +14567 -14177
- package/dist/index.js +270 -56
- package/dist/review-refuter-btpgnkcd.md +77 -0
- package/package.json +2 -2
- package/prompts/default/PROVENANCE.md +1 -0
- package/prompts/default/review-refuter.md +6 -0
- package/skills/crysk/SKILL.md +3 -3
- package/skills/crysk-ci-setup/SKILL.md +4 -4
- package/skills/crysk-ci-setup/references/ci-admission.md +2 -2
- package/skills/crysk-loop/SKILL.md +19 -8
- package/skills/crysk-pr/SKILL.md +1 -1
- package/src/assets.ts +2 -2
- package/src/ci/gates.ts +9 -6
- package/src/ci/outcome.ts +4 -4
- package/src/ci/platform-config.ts +127 -0
- package/src/ci/platform-report.ts +702 -0
- package/src/ci/platform.ts +449 -0
- package/src/ci/reporter.ts +9 -7
- package/src/ci/review-admission.ts +8 -4
- package/src/ci/setup.ts +112 -9
- package/src/commands/ci-setup.ts +21 -18
- package/src/commands/init.ts +21 -12
- package/src/compare/loop.ts +9 -1
- package/src/compare/martian-adapter.ts +1 -1
- package/src/doctor.ts +48 -29
- package/src/index.ts +1 -1
- package/src/menu-context.ts +1 -1
- package/src/pr/admission.ts +53 -11
- package/src/pr/ci-admission-gate.ts +2 -1
- package/src/pr/finding-state-record.ts +58 -0
- package/src/pr/gate-record.ts +22 -0
- package/src/pr/inline.ts +3 -0
- package/src/pr/posting.ts +4 -0
- package/src/pr/pr-status.ts +27 -7
- package/src/pr/pr.ts +39 -8
- package/src/pr/preflight.ts +43 -1
- package/src/pr/publish-outcome.ts +10 -2
- package/src/pr/review-pr.ts +5 -3
- package/src/pr/target.ts +16 -3
- package/src/rereview/classify.ts +4 -0
- package/src/rereview/prepare.ts +47 -22
- package/src/rereview/state.ts +34 -8
- package/src/review/boundary.ts +2 -1
- package/src/review/materiality-gate.ts +25 -1
- package/src/review/pipeline.ts +60 -15
- package/src/review/preflight.ts +132 -28
- package/src/review/prompt-set.ts +12 -8
- package/src/review/report.ts +34 -14
- package/src/review/review.ts +14 -5
- package/src/review/run.ts +54 -18
- package/src/security/published-text.ts +161 -0
- package/src/triage/triage.ts +2 -1
- package/src/updater.ts +1 -1
- package/src/wizard.ts +35 -9
|
@@ -0,0 +1,58 @@
|
|
|
1
|
+
import { writeFileSync } from "node:fs";
|
|
2
|
+
import path from "node:path";
|
|
3
|
+
|
|
4
|
+
export const FINDING_STATE_FILE = "finding-state.json";
|
|
5
|
+
|
|
6
|
+
// Which findings stay open after this run, under the ids that last across runs (R###). findings.json
|
|
7
|
+
// cannot say: it holds only this run's findings, renumbered each run, and a finding missing from it
|
|
8
|
+
// may simply not have been looked at. Only ids and statuses: nothing here can quote the customer's
|
|
9
|
+
// code, so it may leave the runner in metadata-only mode too.
|
|
10
|
+
export interface FindingStateRecord {
|
|
11
|
+
open: { id: string; status: string; tier: string }[];
|
|
12
|
+
// This run's finding id -> its stable id.
|
|
13
|
+
fresh: Record<string, string>;
|
|
14
|
+
// The stable ids this run verified gone. Absence never means resolved.
|
|
15
|
+
resolved: string[];
|
|
16
|
+
// The stable ids a stronger finding replaced this run, under a new id. Not fixed, so not resolved.
|
|
17
|
+
superseded: string[];
|
|
18
|
+
// Rows the state block's size cap left out, whose state is therefore unknown.
|
|
19
|
+
dropped: number;
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
export function buildFindingStateRecord(input: {
|
|
23
|
+
rows: readonly { id: string; status: string; tier: string }[];
|
|
24
|
+
freshFindingIds: readonly string[];
|
|
25
|
+
freshIds: readonly (string | undefined)[];
|
|
26
|
+
resolvedIds: readonly string[];
|
|
27
|
+
supersededIds: readonly string[];
|
|
28
|
+
dropped: number;
|
|
29
|
+
}): FindingStateRecord {
|
|
30
|
+
const fresh: Record<string, string> = {};
|
|
31
|
+
input.freshFindingIds.forEach((findingId, index) => {
|
|
32
|
+
const stable = input.freshIds[index];
|
|
33
|
+
if (stable !== undefined) fresh[findingId] = stable;
|
|
34
|
+
});
|
|
35
|
+
return {
|
|
36
|
+
open: input.rows.map(({ id, status, tier }) => ({ id, status, tier })),
|
|
37
|
+
fresh,
|
|
38
|
+
resolved: [...input.resolvedIds],
|
|
39
|
+
superseded: [...input.supersededIds],
|
|
40
|
+
dropped: input.dropped,
|
|
41
|
+
};
|
|
42
|
+
}
|
|
43
|
+
|
|
44
|
+
// The report to the platform runs as another process, so it reads this file. A failure to write only
|
|
45
|
+
// costs the platform this detail; it must not change the review's result.
|
|
46
|
+
export function writeFindingStateRecord(
|
|
47
|
+
runDir: string,
|
|
48
|
+
record: FindingStateRecord,
|
|
49
|
+
): void {
|
|
50
|
+
try {
|
|
51
|
+
writeFileSync(
|
|
52
|
+
path.join(runDir, FINDING_STATE_FILE),
|
|
53
|
+
`${JSON.stringify(record)}\n`,
|
|
54
|
+
);
|
|
55
|
+
} catch {
|
|
56
|
+
// see above
|
|
57
|
+
}
|
|
58
|
+
}
|
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
import { writeFileSync } from "node:fs";
|
|
2
|
+
import path from "node:path";
|
|
3
|
+
|
|
4
|
+
export const GATE_RECORD_FILE = "gate.json";
|
|
5
|
+
|
|
6
|
+
// The count the `crysk` status is decided on: this run's blocking findings plus earlier ones still
|
|
7
|
+
// open. findings.json holds only this run's, so a same-head re-review there reads zero while the
|
|
8
|
+
// status stays red. The report to the platform runs as another process, so it reads this file. A
|
|
9
|
+
// failure to write only costs the platform the number; it must not change the review's result.
|
|
10
|
+
export function writeGateRecord(
|
|
11
|
+
runDir: string,
|
|
12
|
+
unresolvedBlocking: number,
|
|
13
|
+
): void {
|
|
14
|
+
try {
|
|
15
|
+
writeFileSync(
|
|
16
|
+
path.join(runDir, GATE_RECORD_FILE),
|
|
17
|
+
`${JSON.stringify({ unresolved_blocking: unresolvedBlocking })}\n`,
|
|
18
|
+
);
|
|
19
|
+
} catch {
|
|
20
|
+
// see above
|
|
21
|
+
}
|
|
22
|
+
}
|
package/src/pr/inline.ts
CHANGED
|
@@ -14,6 +14,7 @@
|
|
|
14
14
|
import type { PrHeroFindingRef } from "#compare/compare";
|
|
15
15
|
import { extractAnchor } from "#review/root-cause";
|
|
16
16
|
import { diffRecordPath, splitDiffRecords } from "#review/size-gate";
|
|
17
|
+
import type { FindingStateRecord } from "./finding-state-record";
|
|
17
18
|
import { claimFingerprint, type ParsedFindingMarker } from "./preflight";
|
|
18
19
|
|
|
19
20
|
// The comparison ref plus the proof_refs posting needs to re-anchor an
|
|
@@ -447,6 +448,8 @@ export function buildPostPlan(input: {
|
|
|
447
448
|
// ---------------------------------------------------------------------------
|
|
448
449
|
|
|
449
450
|
export interface InlinePostOutcome {
|
|
451
|
+
// Absent when this post wrote no state block, so nothing can be said about findings across runs.
|
|
452
|
+
findingState?: FindingStateRecord;
|
|
450
453
|
reviewOutcome: "posted" | "demoted";
|
|
451
454
|
reviewFindingCount: number;
|
|
452
455
|
// Always [] this slice (issues #16/#17): findings no longer POST as
|
package/src/pr/posting.ts
CHANGED
|
@@ -21,6 +21,7 @@ import type { CiMode } from "#pr/preflight";
|
|
|
21
21
|
import type { RereviewProvenance } from "#rereview/prepare";
|
|
22
22
|
import type { FindingsDocument } from "#review/findings";
|
|
23
23
|
import { log } from "#ui/primitives";
|
|
24
|
+
import { writeFindingStateRecord } from "./finding-state-record";
|
|
24
25
|
|
|
25
26
|
export interface PostingStageResult {
|
|
26
27
|
posted: InlinePostOutcome | null;
|
|
@@ -91,6 +92,9 @@ export async function postFindingsIfEnabled(input: {
|
|
|
91
92
|
if (posted) {
|
|
92
93
|
input.onDelivered?.();
|
|
93
94
|
await writePostReceipt(input.runDir, input.pr, input.headSha, posted);
|
|
95
|
+
if (posted.findingState !== undefined) {
|
|
96
|
+
writeFindingStateRecord(input.runDir, posted.findingState);
|
|
97
|
+
}
|
|
94
98
|
log(
|
|
95
99
|
`posted: review ${posted.reviewOutcome} (${posted.reviewFindingCount} ` +
|
|
96
100
|
`finding(s)), ${posted.outsideDiffCount} outside diff, ` +
|
package/src/pr/pr-status.ts
CHANGED
|
@@ -24,6 +24,7 @@ import {
|
|
|
24
24
|
findMarkedCommentId,
|
|
25
25
|
IN_FLIGHT_TTL_MS,
|
|
26
26
|
isInFlightCommitStatus,
|
|
27
|
+
isManualReviewHold,
|
|
27
28
|
latestStatusInContext,
|
|
28
29
|
type ParsedFindingMarker,
|
|
29
30
|
parseFindingMarker,
|
|
@@ -573,6 +574,17 @@ function decideNext(
|
|
|
573
574
|
(a) => a.finding.blocking && a.finding.live_on_head,
|
|
574
575
|
);
|
|
575
576
|
if (gate.state === "pending") {
|
|
577
|
+
// Stage 2 #12: a capped rule (the budget used up, or manual_only) held the head instead of carrying a
|
|
578
|
+
// green no review saw. No run will review it on its own, so waiting is wrong, and so is a routine
|
|
579
|
+
// forced review: a human decides between verifying the head and merging it unverified.
|
|
580
|
+
if (
|
|
581
|
+
isManualReviewHold({ state: gate.state, description: gate.description })
|
|
582
|
+
) {
|
|
583
|
+
return {
|
|
584
|
+
next_action: "override-or-fix",
|
|
585
|
+
reason: `The gate holds this head for a manual review: the automatic review budget is used up, or the review policy is manual (ci_review_policy is manual_only), so no review read it and none will on its own. Comment ${commandHint(REVIEW_COMMAND, "")} to verify it, or ${commandHint(OVERRIDE_COMMAND, "")} to merge it unverified.`,
|
|
586
|
+
};
|
|
587
|
+
}
|
|
576
588
|
if (gate.pending_in_flight) {
|
|
577
589
|
return {
|
|
578
590
|
next_action: "wait",
|
|
@@ -600,11 +612,21 @@ function decideNext(
|
|
|
600
612
|
};
|
|
601
613
|
}
|
|
602
614
|
if (gate.state === "success") {
|
|
615
|
+
// PR #472/#475: a finding that lives only in the summary (outside the diff) has no thread, so nothing
|
|
616
|
+
// can be owed on it, but a done that does not name it let the loop merge without reading it. Every
|
|
617
|
+
// done below carries this note.
|
|
618
|
+
const summaryOnly = analyzed.filter(
|
|
619
|
+
(a) => a.finding.comment_id === null,
|
|
620
|
+
).length;
|
|
621
|
+
const summaryNote =
|
|
622
|
+
summaryOnly === 0
|
|
623
|
+
? ""
|
|
624
|
+
: ` ${summaryOnly} finding(s) live only in the summary comment (outside the diff, no thread to reply in); read them there before merging.`;
|
|
603
625
|
// An override is a human's unlock of the head, so it stays done whatever the findings say.
|
|
604
626
|
if (gate.override.active) {
|
|
605
627
|
return {
|
|
606
628
|
next_action: "done",
|
|
607
|
-
reason: `The gate is green through an override by @${gate.override.by ?? "an unknown login"}; nothing is owed on this head
|
|
629
|
+
reason: `The gate is green through an override by @${gate.override.by ?? "an unknown login"}; nothing is owed on this head.${summaryNote}`,
|
|
608
630
|
};
|
|
609
631
|
}
|
|
610
632
|
// A partial review of this head (some reviewers failed) is not a finished review, and the
|
|
@@ -635,7 +657,7 @@ function decideNext(
|
|
|
635
657
|
) {
|
|
636
658
|
return {
|
|
637
659
|
next_action: "done",
|
|
638
|
-
reason: `The gate is green on answered findings no review verified (${gate.description}); the PR is ready to merge, or comment ${commandHint(REVIEW_COMMAND, "")} for another review
|
|
660
|
+
reason: `The gate is green on answered findings no review verified (${gate.description}); the PR is ready to merge, or comment ${commandHint(REVIEW_COMMAND, "")} for another review.${summaryNote}`,
|
|
639
661
|
};
|
|
640
662
|
}
|
|
641
663
|
// #317: no review ran, so the reason must not read as a review that passed.
|
|
@@ -644,14 +666,12 @@ function decideNext(
|
|
|
644
666
|
) {
|
|
645
667
|
return {
|
|
646
668
|
next_action: "done",
|
|
647
|
-
reason:
|
|
648
|
-
"Every changed file on this head is excluded from review, so there was nothing to review; nothing is owed.",
|
|
669
|
+
reason: `Every changed file on this head is excluded from review, so there was nothing to review; nothing is owed.${summaryNote}`,
|
|
649
670
|
};
|
|
650
671
|
}
|
|
651
672
|
return {
|
|
652
673
|
next_action: "done",
|
|
653
|
-
reason:
|
|
654
|
-
"The review on this head passed the gate, and every finding on this head has a reply.",
|
|
674
|
+
reason: `The review on this head passed the gate, and every finding on this head has a reply.${summaryNote}`,
|
|
655
675
|
};
|
|
656
676
|
}
|
|
657
677
|
// failure: the blocking findings on this head decide between triage, fix, and override.
|
|
@@ -719,7 +739,7 @@ function decideNext(
|
|
|
719
739
|
next_action: "fix",
|
|
720
740
|
reason:
|
|
721
741
|
open === 0 && unseen === 0
|
|
722
|
-
? "Every live blocking finding is answered and a review rule skipped this head; push the fix. If a rule skips that push too, the gate reads the replies on it and can turn green, marked unverified; otherwise the review verifies it."
|
|
742
|
+
? "Every live blocking finding is answered and a review rule skipped this head; push the fix. If a rule skips that push too, the gate reads the replies on it and can turn green, marked unverified, or, under a used-up budget or manual_only, hold the head for a manual review; otherwise the review verifies it."
|
|
723
743
|
: `A live blocking finding is tagged applied, but ${open + unseen} blocking finding(s) are still open (deferred, misclassified, a dismissal not upheld, a reply at another head, or one this read cannot see), so replies alone will not unlock the carried red; push the fix, and answer or fix the rest.`,
|
|
724
744
|
};
|
|
725
745
|
}
|
package/src/pr/pr.ts
CHANGED
|
@@ -74,6 +74,7 @@ import {
|
|
|
74
74
|
WORKFLOW_FILE_NAME,
|
|
75
75
|
} from "../brand";
|
|
76
76
|
import { CliError } from "../errors";
|
|
77
|
+
import { buildFindingStateRecord } from "./finding-state-record";
|
|
77
78
|
import {
|
|
78
79
|
buildPostPlan,
|
|
79
80
|
computeDroppedFindingIds,
|
|
@@ -1310,7 +1311,7 @@ const PR_ISSUE_COMMENTS_QUERY =
|
|
|
1310
1311
|
"repository(owner:$repoOwner,name:$repoName){pullRequest(number:$number){" +
|
|
1311
1312
|
"comments(first:100,after:$cursor){" +
|
|
1312
1313
|
"pageInfo{endCursor hasNextPage}" +
|
|
1313
|
-
"nodes{databaseId body createdAt updatedAt author{login}}}}}}";
|
|
1314
|
+
"nodes{databaseId body createdAt updatedAt author{login __typename}}}}}}";
|
|
1314
1315
|
|
|
1315
1316
|
const PR_REVIEW_COMMENTS_QUERY =
|
|
1316
1317
|
"query($repoOwner:String!,$repoName:String!,$number:Int!,$cursor:String){" +
|
|
@@ -1609,7 +1610,7 @@ function issueCommentsFromNodes(nodes: unknown[]): {
|
|
|
1609
1610
|
if (id === null) continue;
|
|
1610
1611
|
comments.push({
|
|
1611
1612
|
id,
|
|
1612
|
-
user:
|
|
1613
|
+
user: graphqlReviewLogin(record.author),
|
|
1613
1614
|
body: typeof record.body === "string" ? record.body : "",
|
|
1614
1615
|
...(typeof record.createdAt === "string"
|
|
1615
1616
|
? { created_at: record.createdAt }
|
|
@@ -3133,31 +3134,48 @@ export async function postInlineFindings(input: {
|
|
|
3133
3134
|
// WHY one merge function for both writes: the S1 re-merge reruns it with a
|
|
3134
3135
|
// PATCH-sized budget and the SAME inputs, so the kept ids and hw cannot
|
|
3135
3136
|
// drift between the POST and the PATCH.
|
|
3137
|
+
// WHY a first review writes the block too: an outside-diff finding has no
|
|
3138
|
+
// comment of its own, so the block is its only machine-readable copy. A
|
|
3139
|
+
// first review without one left `crysk status` blind to those findings
|
|
3140
|
+
// (PR #472: four advisories, status read none).
|
|
3141
|
+
// WHY not every post without framing (PR #475 R001): over another head's
|
|
3142
|
+
// summary, a block holding only this run's findings would replace the
|
|
3143
|
+
// marker recovery the next review relies on, so still-open findings of
|
|
3144
|
+
// older heads would stop being priors. Such a post writes the block only
|
|
3145
|
+
// as a same-head re-post, which carries the previous rows. A re-review
|
|
3146
|
+
// whose prior summary vanished has its framing dropped and stays
|
|
3147
|
+
// state-less: that framing names threads that no longer exist.
|
|
3148
|
+
const stateFramed =
|
|
3149
|
+
framing !== undefined ||
|
|
3150
|
+
(input.rereview === undefined &&
|
|
3151
|
+
(existingSummaryId === null || previousState?.headSha === doc.head_sha));
|
|
3136
3152
|
const mergeStateWithBudget = (
|
|
3137
3153
|
stateBudget: number,
|
|
3138
3154
|
): MergedState | undefined =>
|
|
3139
|
-
|
|
3155
|
+
!stateFramed
|
|
3140
3156
|
? undefined
|
|
3141
3157
|
: mergeStateFindings({
|
|
3142
3158
|
headSha: doc.head_sha,
|
|
3143
3159
|
reviews: postedReviewCount,
|
|
3144
|
-
live: framing
|
|
3160
|
+
live: framing?.live ?? [],
|
|
3145
3161
|
previous: previousState,
|
|
3146
3162
|
previousHead: previousState?.headSha ?? null,
|
|
3147
3163
|
retiredIds: [
|
|
3148
|
-
...(framing
|
|
3149
|
-
...(framing
|
|
3164
|
+
...(framing?.resolved_ids ?? []),
|
|
3165
|
+
...(framing?.worsened ?? []).map((w) => w.priorId),
|
|
3150
3166
|
],
|
|
3151
3167
|
// The priors this run's gate count released through a materiality
|
|
3152
3168
|
// demotion, recorded on the rereview block by the pipeline, so the
|
|
3153
3169
|
// state the next run reads agrees with this run's count.
|
|
3154
|
-
...(
|
|
3170
|
+
...(framing?.materiality_released_ids === undefined
|
|
3155
3171
|
? {}
|
|
3156
|
-
: { releasedIds:
|
|
3172
|
+
: { releasedIds: framing.materiality_released_ids }),
|
|
3157
3173
|
fresh: freshFindings,
|
|
3158
3174
|
stateBudget,
|
|
3159
3175
|
});
|
|
3160
3176
|
const merged = mergeStateWithBudget(stateBlockBudget(referenceBody.length));
|
|
3177
|
+
// What was finally written, for the record the platform reads: a re-merge below may drop rows.
|
|
3178
|
+
let writtenState = merged;
|
|
3161
3179
|
|
|
3162
3180
|
const renderWithState = (
|
|
3163
3181
|
delta: PrCommentDelta,
|
|
@@ -3361,6 +3379,7 @@ export async function postInlineFindings(input: {
|
|
|
3361
3379
|
stateBlockBudget(bodyWithoutState.length),
|
|
3362
3380
|
);
|
|
3363
3381
|
patchBody = renderPatch(reMerged, 0);
|
|
3382
|
+
writtenState = reMerged;
|
|
3364
3383
|
}
|
|
3365
3384
|
if (patchBody.length > GITHUB_ISSUE_COMMENT_MAX) {
|
|
3366
3385
|
process.stdout.write(
|
|
@@ -3515,6 +3534,18 @@ export async function postInlineFindings(input: {
|
|
|
3515
3534
|
commentUrls: commentUrlByFindingId,
|
|
3516
3535
|
movedHeadSha,
|
|
3517
3536
|
...(closingCommentId === undefined ? {} : { closingCommentId }),
|
|
3537
|
+
...(writtenState === undefined
|
|
3538
|
+
? {}
|
|
3539
|
+
: {
|
|
3540
|
+
findingState: buildFindingStateRecord({
|
|
3541
|
+
rows: writtenState.rows,
|
|
3542
|
+
freshFindingIds: postedDoc.findings.map((f) => f.id),
|
|
3543
|
+
freshIds: writtenState.freshIds,
|
|
3544
|
+
resolvedIds: framing?.resolved_ids ?? [],
|
|
3545
|
+
supersededIds: (framing?.worsened ?? []).map((w) => w.priorId),
|
|
3546
|
+
dropped: writtenState.dropped,
|
|
3547
|
+
}),
|
|
3548
|
+
}),
|
|
3518
3549
|
};
|
|
3519
3550
|
}
|
|
3520
3551
|
|
package/src/pr/preflight.ts
CHANGED
|
@@ -1007,10 +1007,51 @@ export function latestStatusInContext(
|
|
|
1007
1007
|
return latest;
|
|
1008
1008
|
}
|
|
1009
1009
|
|
|
1010
|
+
// The pending a gate skip writes when a rule that caps automatic reviews (the budget is used up, or the
|
|
1011
|
+
// policy is manual_only) would otherwise carry a green no review saw (ROADMAP Stage 2 #12). It holds the
|
|
1012
|
+
// head until a human asks for a review or an override. `label` is the rule (ciReviewSkipLabel); `sha` is
|
|
1013
|
+
// the commit whose verdict it held, so the next skip's walk can jump there (heldVerdictSha).
|
|
1014
|
+
export function manualReviewHoldDescription(
|
|
1015
|
+
label: string,
|
|
1016
|
+
sha: string,
|
|
1017
|
+
): string {
|
|
1018
|
+
return clipStatusDescription(
|
|
1019
|
+
`${STATUS_CONTEXT}: waiting for a manual review (${label}); carried from ${sha.slice(0, 7)} — /pr-hero review or /pr-hero override`,
|
|
1020
|
+
);
|
|
1021
|
+
}
|
|
1022
|
+
|
|
1023
|
+
const MANUAL_REVIEW_HOLD = /^crysk: waiting for a manual review \(/;
|
|
1024
|
+
const HELD_VERDICT =
|
|
1025
|
+
/^crysk: waiting for a manual review \((.+)\); carried from ([0-9a-f]{7}) — /;
|
|
1026
|
+
|
|
1027
|
+
// Whether a pending status is the gate's hold, not a running review. No run holds it, so no in-flight
|
|
1028
|
+
// reader may wait on it: it lasts until a human acts, not until a run ends.
|
|
1029
|
+
export function isManualReviewHold(
|
|
1030
|
+
status: Pick<CommitStatusFact, "state" | "description">,
|
|
1031
|
+
): boolean {
|
|
1032
|
+
return (
|
|
1033
|
+
status.state === "pending" &&
|
|
1034
|
+
MANUAL_REVIEW_HOLD.test(status.description ?? "")
|
|
1035
|
+
);
|
|
1036
|
+
}
|
|
1037
|
+
|
|
1038
|
+
// The short sha a hold names, or null. WHY the text is rebuilt: the builder above is the only authority on
|
|
1039
|
+
// a hold's exact text, so a description that only resembles one never redirects the carry's walk.
|
|
1040
|
+
export function heldVerdictSha(
|
|
1041
|
+
status: Pick<CommitStatusFact, "state" | "description">,
|
|
1042
|
+
): string | null {
|
|
1043
|
+
if (!isManualReviewHold(status)) return null;
|
|
1044
|
+
const description = status.description ?? "";
|
|
1045
|
+
const match = HELD_VERDICT.exec(description);
|
|
1046
|
+
if (match === null) return null;
|
|
1047
|
+
const [, label, sha] = match as unknown as [string, string, string];
|
|
1048
|
+
return manualReviewHoldDescription(label, sha) === description ? sha : null;
|
|
1049
|
+
}
|
|
1050
|
+
|
|
1010
1051
|
// Latest status for our context, by created_at — GitHub's list is usually
|
|
1011
1052
|
// newest-first, but the watcher must not depend on that. A pending older
|
|
1012
1053
|
// than the TTL is NOT in-flight: the process likely died, and the next tick
|
|
1013
|
-
// must be allowed to launch.
|
|
1054
|
+
// must be allowed to launch. A manual-review hold is never in flight.
|
|
1014
1055
|
// In-flight in ANY of the given contexts: each context's own latest status is
|
|
1015
1056
|
// judged, so a fresh pending on one never hides behind a settled one on another.
|
|
1016
1057
|
// The default reads the gate alone, which is a `--ci` run's yield. A non-CI run
|
|
@@ -1024,6 +1065,7 @@ export function isInFlightCommitStatus(
|
|
|
1024
1065
|
return contexts.some((context) => {
|
|
1025
1066
|
const latest = latestStatusInContext(statuses, context);
|
|
1026
1067
|
if (latest === null || latest.state !== "pending") return false;
|
|
1068
|
+
if (isManualReviewHold(latest)) return false;
|
|
1027
1069
|
const createdMs = Date.parse(latest.created_at);
|
|
1028
1070
|
if (Number.isNaN(createdMs)) return false;
|
|
1029
1071
|
return nowMs - createdMs < ttlMs;
|
|
@@ -105,7 +105,9 @@ export interface PublishRunOutcomeInput {
|
|
|
105
105
|
// (settleCommitStatusAndLedger) reads this count in reviewPr()'s finally,
|
|
106
106
|
// and a throw later in this tail must not leave it unknown. The count is the
|
|
107
107
|
// gate's UNRESOLVED blocking count (rereview/state.ts), not this run's survivors.
|
|
108
|
-
|
|
108
|
+
// `conclusive` is false for a zero the gate itself ignores (commitStatusCompletion),
|
|
109
|
+
// because a reviewer that never ran or failed might have been the one to block.
|
|
110
|
+
onBlockingCount: (count: number, conclusive: boolean) => void;
|
|
109
111
|
// Test-only seam (default: the real postFindingsIfEnabled). The posting stage
|
|
110
112
|
// reaches gh with no offline seam of its own, so the suite passes a fake that
|
|
111
113
|
// rejects or resolves, the way it drives the throw and the delivery point.
|
|
@@ -153,7 +155,13 @@ export async function publishRunOutcome(
|
|
|
153
155
|
doc.findings,
|
|
154
156
|
result.gatePriors ?? [],
|
|
155
157
|
);
|
|
156
|
-
input.onBlockingCount(
|
|
158
|
+
input.onBlockingCount(
|
|
159
|
+
blockingCount,
|
|
160
|
+
blockingCount > 0 ||
|
|
161
|
+
(result.sessionFailed !== true &&
|
|
162
|
+
result.noReviewerRan !== true &&
|
|
163
|
+
(result.hunterFailures?.length ?? 0) === 0),
|
|
164
|
+
);
|
|
157
165
|
const reportPath = await writeRunReport({
|
|
158
166
|
runDir: input.runDir,
|
|
159
167
|
doc,
|
package/src/pr/review-pr.ts
CHANGED
|
@@ -92,7 +92,7 @@ import {
|
|
|
92
92
|
resolveMaxConcurrentSteps,
|
|
93
93
|
resolveSharedPrefix,
|
|
94
94
|
} from "#review/preflight";
|
|
95
|
-
import { estimateCost } from "#review/report";
|
|
95
|
+
import { code, estimateCost } from "#review/report";
|
|
96
96
|
import {
|
|
97
97
|
pipelineConsolidateInput,
|
|
98
98
|
pipelineMaterialityInput,
|
|
@@ -121,6 +121,7 @@ import { acquirePidLock } from "../home";
|
|
|
121
121
|
import { worktreeLockPath } from "../home-preflight";
|
|
122
122
|
import { readLocalIgnoreRules } from "../ignore-read";
|
|
123
123
|
import type { ProductionRuntime } from "../production-runtime";
|
|
124
|
+
import { writeGateRecord } from "./gate-record";
|
|
124
125
|
|
|
125
126
|
// The CLI's record for a pipeline whose every hunter failed (cli.json,
|
|
126
127
|
// odd/tasks/agent-first-ci.md, T2c). reviewPr writes it right after the pipeline's
|
|
@@ -228,7 +229,7 @@ export async function postAdjudicationReplies(input: {
|
|
|
228
229
|
});
|
|
229
230
|
const evidence =
|
|
230
231
|
entry.evidence.length > 0
|
|
231
|
-
? `\n\nEvidence:\n${entry.evidence.map((e) => `-
|
|
232
|
+
? `\n\nEvidence:\n${entry.evidence.map((e) => `- ${code(e)}`).join("\n")}`
|
|
232
233
|
: "";
|
|
233
234
|
const body = `${marker}\n\n${ADJUDICATION_BADGES[entry.verdict]}${evidence}`;
|
|
234
235
|
let commentId: number | undefined;
|
|
@@ -916,8 +917,9 @@ export async function reviewPr(
|
|
|
916
917
|
onPosted: (outcome) => {
|
|
917
918
|
posted = outcome;
|
|
918
919
|
},
|
|
919
|
-
onBlockingCount: (count) => {
|
|
920
|
+
onBlockingCount: (count, conclusive) => {
|
|
920
921
|
blockingCount = count;
|
|
922
|
+
if (conclusive) writeGateRecord(runDir, count);
|
|
921
923
|
},
|
|
922
924
|
});
|
|
923
925
|
// T5b: the verdicts are posted only after the review itself published.
|
package/src/pr/target.ts
CHANGED
|
@@ -49,8 +49,10 @@ import type { estimateCost } from "#review/report";
|
|
|
49
49
|
import {
|
|
50
50
|
type LoadedRunConfig,
|
|
51
51
|
loadRunConfig,
|
|
52
|
-
|
|
52
|
+
NO_INSTRUCTIONS_NOTICE,
|
|
53
53
|
resolvePromptSet,
|
|
54
|
+
resolveReviewInstructions,
|
|
55
|
+
shadowedGotchasNotice,
|
|
54
56
|
} from "#review/run";
|
|
55
57
|
import type { SizeGateConfig } from "#review/size-gate";
|
|
56
58
|
import type { ReviewSpec } from "#review/spec";
|
|
@@ -142,7 +144,8 @@ export interface ResolvedPrPromptSet {
|
|
|
142
144
|
spec: ReviewSpec;
|
|
143
145
|
agentFiles: Map<string, ParsedAgent>;
|
|
144
146
|
promptSet: PromptSetIdentity;
|
|
145
|
-
|
|
147
|
+
// Undefined when the repo has no REVIEW.md or gotchas file (S3).
|
|
148
|
+
gotchasPath: string | undefined;
|
|
146
149
|
}
|
|
147
150
|
|
|
148
151
|
// Step 1's second half: the CI budget ceiling (+ its notices), then the
|
|
@@ -216,10 +219,20 @@ export async function resolvePrPromptSetAndBudget(params: {
|
|
|
216
219
|
params.options,
|
|
217
220
|
params.loaded,
|
|
218
221
|
);
|
|
219
|
-
const
|
|
222
|
+
const instructions = resolveReviewInstructions(
|
|
220
223
|
params.options.gotchas,
|
|
221
224
|
params.operatorRoot,
|
|
222
225
|
);
|
|
226
|
+
const notice =
|
|
227
|
+
instructions.shadowed !== undefined
|
|
228
|
+
? shadowedGotchasNotice(instructions.shadowed)
|
|
229
|
+
: instructions.path === undefined
|
|
230
|
+
? NO_INSTRUCTIONS_NOTICE
|
|
231
|
+
: undefined;
|
|
232
|
+
if (notice !== undefined) {
|
|
233
|
+
log(params.isCi ? formatWorkflowCommand("notice", notice) : notice);
|
|
234
|
+
}
|
|
235
|
+
const gotchasPath = instructions.path;
|
|
223
236
|
return {
|
|
224
237
|
ciBudgetCeiling,
|
|
225
238
|
ciSizeLimits,
|
package/src/rereview/classify.ts
CHANGED
|
@@ -107,6 +107,10 @@ export interface PriorRecord {
|
|
|
107
107
|
channel: FindingChannel;
|
|
108
108
|
locs: readonly string[];
|
|
109
109
|
claim: string;
|
|
110
|
+
// The fingerprint of the claim as first written, which is what the posted
|
|
111
|
+
// marker carries. `claim` here may be its published form, so hashing it would
|
|
112
|
+
// not match. Absent when the prior has no claim to fingerprint.
|
|
113
|
+
c?: string;
|
|
110
114
|
triage: PriorTriage | null;
|
|
111
115
|
// Case B rule 7b: any new reply on the prior thread, including tags
|
|
112
116
|
// other than a newly-applied `applied` (that one is rule 7).
|
package/src/rereview/prepare.ts
CHANGED
|
@@ -4,7 +4,7 @@
|
|
|
4
4
|
// The case machine itself stays in rereview/plan.ts.
|
|
5
5
|
|
|
6
6
|
import { normalizePath } from "#compare/compare";
|
|
7
|
-
import { claimFingerprint } from "#pr/preflight";
|
|
7
|
+
import { claimFingerprint, parseFindingMarker } from "#pr/preflight";
|
|
8
8
|
import type { Severity } from "#review/findings";
|
|
9
9
|
import { parseFindingCommentBadge } from "#review/report";
|
|
10
10
|
import {
|
|
@@ -617,20 +617,20 @@ export function priorsFromStateFindings(
|
|
|
617
617
|
channel: finding.channel,
|
|
618
618
|
locs: finding.locs,
|
|
619
619
|
claim: finding.claim,
|
|
620
|
+
c: finding.c,
|
|
620
621
|
triage: null,
|
|
621
622
|
newThreadReply: false,
|
|
622
623
|
}));
|
|
623
624
|
}
|
|
624
625
|
|
|
625
626
|
// #206's fix: the fallback that runs whenever the previous summary's
|
|
626
|
-
// `pr-hero-state` block cannot supply priors — missing
|
|
627
|
-
//
|
|
628
|
-
//
|
|
629
|
-
//
|
|
630
|
-
//
|
|
631
|
-
//
|
|
632
|
-
//
|
|
633
|
-
// `claimFingerprint("")`'s shared constant.
|
|
627
|
+
// `pr-hero-state` block cannot supply priors — missing or unparseable. Before
|
|
628
|
+
// 2026-11-04 a FIRST review never wrote a state block, so this was the
|
|
629
|
+
// routine path of every second review; summaries posted before then still
|
|
630
|
+
// reach it. It used to hardcode `sev: "WARNING"` / `tier: "advisory"` /
|
|
631
|
+
// `claim: ""` for EVERY prior here — silently downgrading a live BLOCKER to a
|
|
632
|
+
// warning on every second review of every PR, and collapsing every claim-less
|
|
633
|
+
// row onto `claimFingerprint("")`'s shared constant.
|
|
634
634
|
//
|
|
635
635
|
// `body` is the posted comment's raw text (the marker line alone carries no
|
|
636
636
|
// sev/tier/claim — `findingMarker` signs only the claim's fingerprint, never
|
|
@@ -651,6 +651,7 @@ export function priorsFromPostedMarkers(
|
|
|
651
651
|
): PriorRecord[] {
|
|
652
652
|
return posted.map((item, i) => {
|
|
653
653
|
const recovered = parseFindingCommentBadge(item.body);
|
|
654
|
+
const c = recovered === null ? undefined : parseFindingMarker(item.body)?.c;
|
|
654
655
|
return {
|
|
655
656
|
id: `R${String(i + 1).padStart(3, "0")}`,
|
|
656
657
|
sev: recovered?.sev ?? SEVERITY_UNRECOVERABLE,
|
|
@@ -658,6 +659,7 @@ export function priorsFromPostedMarkers(
|
|
|
658
659
|
channel: item.channel,
|
|
659
660
|
locs: [`${item.path}:${item.line}`],
|
|
660
661
|
claim: recovered?.claim ?? "",
|
|
662
|
+
...(c === undefined ? {} : { c }),
|
|
661
663
|
triage: null,
|
|
662
664
|
newThreadReply: false,
|
|
663
665
|
};
|
|
@@ -750,13 +752,18 @@ export interface PostedForPrior {
|
|
|
750
752
|
// triage tag or a thread left open; an over-match puts a "✅ RESOLVED ·
|
|
751
753
|
// verified gone" reply on a still-live finding.
|
|
752
754
|
export function bindPriorsToPosted<T extends PostedForPrior>(
|
|
753
|
-
priors: readonly {
|
|
755
|
+
priors: readonly {
|
|
756
|
+
id: string;
|
|
757
|
+
locs: readonly string[];
|
|
758
|
+
claim: string;
|
|
759
|
+
c?: string;
|
|
760
|
+
}[],
|
|
754
761
|
posted: readonly T[],
|
|
755
762
|
window: number = IDENTITY_LINE_WINDOW,
|
|
756
763
|
): Map<string, T> {
|
|
757
764
|
interface Pair {
|
|
758
765
|
priorId: string;
|
|
759
|
-
|
|
766
|
+
fingerprint: string | undefined;
|
|
760
767
|
row: T;
|
|
761
768
|
distance: number;
|
|
762
769
|
}
|
|
@@ -767,7 +774,12 @@ export function bindPriorsToPosted<T extends PostedForPrior>(
|
|
|
767
774
|
for (const row of posted) {
|
|
768
775
|
const distance = anchorDistance(identity, row, window);
|
|
769
776
|
if (distance === null) continue;
|
|
770
|
-
pairs.push({
|
|
777
|
+
pairs.push({
|
|
778
|
+
priorId: prior.id,
|
|
779
|
+
fingerprint: priorFingerprint(prior),
|
|
780
|
+
row,
|
|
781
|
+
distance,
|
|
782
|
+
});
|
|
771
783
|
}
|
|
772
784
|
}
|
|
773
785
|
|
|
@@ -804,7 +816,7 @@ export function bindPriorsToPosted<T extends PostedForPrior>(
|
|
|
804
816
|
let changed = false;
|
|
805
817
|
const proposals: Pair[] = [];
|
|
806
818
|
for (const [priorId, group] of byPrior) {
|
|
807
|
-
const winner = resolvePriorTie(group, group[0]?.
|
|
819
|
+
const winner = resolvePriorTie(group, group[0]?.fingerprint);
|
|
808
820
|
if (winner === undefined) {
|
|
809
821
|
disqualified.add(priorId);
|
|
810
822
|
changed = true;
|
|
@@ -869,16 +881,25 @@ function anchorDistance(
|
|
|
869
881
|
return best;
|
|
870
882
|
}
|
|
871
883
|
|
|
884
|
+
// What a prior's marker should carry. No claim, no fingerprint:
|
|
885
|
+
// `claimFingerprint("")` is a real constant hash (`e3b0c44298fc`) that every
|
|
886
|
+
// empty-claim prior shares, so consulting it would hand a whole tied set to
|
|
887
|
+
// whichever row came first. A stored `c` wins because the stored claim may be
|
|
888
|
+
// its published form, which the posted marker never hashed.
|
|
889
|
+
function priorFingerprint(prior: {
|
|
890
|
+
claim: string;
|
|
891
|
+
c?: string;
|
|
892
|
+
}): string | undefined {
|
|
893
|
+
if (prior.claim.trim().length === 0) return undefined;
|
|
894
|
+
return prior.c ?? claimFingerprint(prior.claim);
|
|
895
|
+
}
|
|
896
|
+
|
|
872
897
|
function resolvePriorTie<T extends PostedForPrior>(
|
|
873
898
|
tied: readonly { row: T; distance: number }[],
|
|
874
|
-
|
|
899
|
+
fingerprint: string | undefined,
|
|
875
900
|
): T | undefined {
|
|
876
901
|
if (tied.length === 1) return tied[0]?.row;
|
|
877
|
-
|
|
878
|
-
// (`e3b0c44298fc`) that every empty-claim prior shares, so consulting it
|
|
879
|
-
// here would hand the whole tied set to whichever row came first.
|
|
880
|
-
if (claim.trim().length === 0) return undefined;
|
|
881
|
-
const fingerprint = claimFingerprint(claim);
|
|
902
|
+
if (fingerprint === undefined) return undefined;
|
|
882
903
|
const matching = tied.filter((c) => c.row.marker.c === fingerprint);
|
|
883
904
|
return matching.length === 1 ? matching[0]?.row : undefined;
|
|
884
905
|
}
|
|
@@ -890,13 +911,17 @@ function resolvePriorTie<T extends PostedForPrior>(
|
|
|
890
911
|
// claim-less prior shares), and anything short of exactly one match leaves the
|
|
891
912
|
// comment to nobody.
|
|
892
913
|
function resolveRowTie<T extends PostedForPrior>(
|
|
893
|
-
contenders: readonly {
|
|
914
|
+
contenders: readonly {
|
|
915
|
+
priorId: string;
|
|
916
|
+
fingerprint: string | undefined;
|
|
917
|
+
row: T;
|
|
918
|
+
}[],
|
|
894
919
|
): { priorId: string; row: T } | undefined {
|
|
895
920
|
if (contenders.length === 1) return contenders[0];
|
|
896
921
|
const matching = contenders.filter(
|
|
897
922
|
(contender) =>
|
|
898
|
-
contender.
|
|
899
|
-
contender.row.marker.c ===
|
|
923
|
+
contender.fingerprint !== undefined &&
|
|
924
|
+
contender.row.marker.c === contender.fingerprint,
|
|
900
925
|
);
|
|
901
926
|
return matching.length === 1 ? matching[0] : undefined;
|
|
902
927
|
}
|