crysk 0.6.0 → 0.6.1

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.
@@ -27,7 +27,6 @@
27
27
  import path from "node:path";
28
28
  import { classifyCliError } from "#ci/classify-error";
29
29
  import { recordReviewDelivered, writeCliRecord } from "#ci/cli-record";
30
- import { ciExitCode } from "#ci/gates";
31
30
  import {
32
31
  ingestReviewMetrics,
33
32
  notionalCostInput,
@@ -42,11 +41,12 @@ import {
42
41
  } from "#pr/admission";
43
42
  import { publishCiReviewIfEligible } from "#pr/ci-publish";
44
43
  import type { PrDiscoveryResult } from "#pr/discovery";
45
- import { type InlinePostOutcome, postingExitCode } from "#pr/inline";
44
+ import type { InlinePostOutcome } from "#pr/inline";
46
45
  import type { PostMinSeverity } from "#pr/post-floor";
47
46
  import { type PostingStageResult, postFindingsIfEnabled } from "#pr/posting";
48
47
  import type { CiMode } from "#pr/preflight";
49
48
  import { unresolvedBlockingCount } from "#rereview/state";
49
+ import { reviewExitCode } from "#review/exit-code";
50
50
  import type { Telemetry } from "#review/findings";
51
51
  import type { PipelineResult } from "#review/pipeline";
52
52
  import type { CliOptions } from "#review/preflight";
@@ -359,15 +359,16 @@ export async function publishRunOutcome(
359
359
  "completed",
360
360
  "review complete",
361
361
  );
362
- // Assistant posture (spec 2.1): in CI mode, exit 0 even with blocking
363
- // findings — ciExitCode only fails on a fatal session failure (already
364
- // returned above) or a genuine posting drop (design D6). Outside CI,
365
- // postingExitCode keeps its existing behavior unchanged.
366
- return input.isCi
367
- ? ciExitCode({
368
- sessionFailed: result.sessionFailed,
369
- droppedFindingIds: posted?.droppedFindingIds.length ?? 0,
370
- blockingCount,
371
- })
372
- : postingExitCode(posted);
362
+ // One table for every mode (src/review/exit-code.ts): findings never change
363
+ // the code (spec 2.1), an incomplete review exits 3, a review that did not
364
+ // happen or a dropped posting exits 1.
365
+ return reviewExitCode({
366
+ sessionFailed: result.sessionFailed,
367
+ ...(result.noReviewerRan === undefined
368
+ ? {}
369
+ : { noReviewerRan: result.noReviewerRan }),
370
+ runStatus: doc.run_status,
371
+ droppedFindingIds: posted?.droppedFindingIds.length ?? 0,
372
+ blockingCount,
373
+ });
373
374
  }
@@ -114,6 +114,7 @@ import {
114
114
  type AdjudicationOutcome,
115
115
  adjudicationMarker,
116
116
  CI_ADJUDICATOR_LOGIN,
117
+ trustedAdjudicator,
117
118
  } from "#triage/adjudication";
118
119
  import { dryRunHunterCount, reviewingLine } from "#ui/plan";
119
120
  import { log, styleEnabled, terminalWidth } from "#ui/primitives";
@@ -264,9 +265,23 @@ export async function postAdjudicationReplies(input: {
264
265
  commentId,
265
266
  spawnFn: input.spawnFn,
266
267
  }).catch(() => undefined);
267
- if (author !== CI_ADJUDICATOR_LOGIN) {
268
+ // In platform mode the run posts as the platform App's bot, the login the platform's token
269
+ // exchange returned (CRYSK_PLATFORM_ACTOR). Its ruling counts only on a finding that bot
270
+ // posted, so the finding's own author is read too, as the gate's read does.
271
+ const platformActor = process.env.CRYSK_PLATFORM_ACTOR;
272
+ const findingAuthor =
273
+ author === undefined || author === CI_ADJUDICATOR_LOGIN
274
+ ? undefined
275
+ : await readCommentAuthor({
276
+ operatorRoot: input.operatorRoot,
277
+ commentId: entry.threadId,
278
+ spawnFn: input.spawnFn,
279
+ }).catch(() => undefined);
280
+ if (
281
+ !trustedAdjudicator(author, findingAuthor, { kind: "ci", platformActor })
282
+ ) {
268
283
  process.stdout.write(
269
- `::warning::crysk: the CI adjudication for ${entry.priorId} was not recorded under ${CI_ADJUDICATOR_LOGIN}; the finding still counts toward the gate\n`,
284
+ `::warning::crysk: the CI adjudication for ${entry.priorId} was not recorded under ${CI_ADJUDICATOR_LOGIN} or the platform's bot; the finding still counts toward the gate\n`,
270
285
  );
271
286
  }
272
287
  }
@@ -9,9 +9,9 @@ import type { Severity } from "#review/findings";
9
9
  import { parseFindingCommentBadge } from "#review/report";
10
10
  import {
11
11
  type AdjudicationSubject,
12
- CI_ADJUDICATOR_LOGIN,
13
12
  MAX_ADJUDICATIONS_PER_RUN,
14
13
  parseAdjudicationMarker,
14
+ trustedAdjudicator,
15
15
  } from "#triage/adjudication";
16
16
  import { parseTriageMarker } from "#triage/triage";
17
17
  import {
@@ -981,6 +981,8 @@ export function enrichPriorsFromThreads(input: {
981
981
  created_at?: string;
982
982
  }[];
983
983
  summaryUpdatedAt: string | null;
984
+ // In platform mode, the login the platform's token exchange returned (CRYSK_PLATFORM_ACTOR).
985
+ platformActor?: string;
984
986
  }): PriorRecord[] {
985
987
  const bound = bindPriorsToPosted(input.priors, input.posted);
986
988
  return input.priors.map((prior) => {
@@ -1010,8 +1012,17 @@ export function enrichPriorsFromThreads(input: {
1010
1012
  // so a verdict on an earlier head is stale and a later head needs a new
1011
1013
  // recheck. With no headSha, no bot verdict can be shown current.
1012
1014
  if (triage?.replyId !== undefined && input.headSha !== undefined) {
1015
+ // The finding itself sits among the replies (a review comment with no parent).
1016
+ const findingAuthor = input.replies.find((r) => r.id === parent.id)?.user;
1013
1017
  for (const reply of thread) {
1014
- if (reply.user !== CI_ADJUDICATOR_LOGIN) continue;
1018
+ if (
1019
+ !trustedAdjudicator(reply.user, findingAuthor, {
1020
+ kind: "ci",
1021
+ platformActor: input.platformActor,
1022
+ })
1023
+ ) {
1024
+ continue;
1025
+ }
1015
1026
  const adjudication = parseAdjudicationMarker(reply.body);
1016
1027
  if (
1017
1028
  adjudication === null ||
@@ -0,0 +1,60 @@
1
+ // The exit code of a review, in every mode: local, `--pr`, and headless CI.
2
+ //
3
+ // The exit code says whether the TOOL did its job, never what it found.
4
+ // Spec 2.1, the assistant posture: "The CLI MUST NOT exit with a non-zero code
5
+ // merely because findings (even `blocking` ones) were discovered". The
6
+ // verdict belongs to the `crysk` commit status and the posted comments.
7
+ //
8
+ // 0 the review is complete, with or without findings
9
+ // 1 no usable review: every hunter failed, no reviewer ran, or findings that
10
+ // should have been posted were dropped
11
+ // 2 usage error (src/cli.ts, before any review starts)
12
+ // 3 the review is incomplete: it ran, but a step failed (a hunter, the
13
+ // refuter, verify, adjudicate, or the pipeline ceiling). Run it again.
14
+ //
15
+ // WHY 3 exists, instead of folding an incomplete review into 0 or 1: until
16
+ // this module, an incomplete review exited 0. In GitHub the job check then
17
+ // stayed green while the gate's own `crysk` status read "review partial", and
18
+ // a pipeline outside GitHub (a script running `crysk review && merge`) had no
19
+ // machine-readable way to learn that a reviewer never looked. Exit 1 would
20
+ // have fixed the signal but merged two different next steps: 1 means
21
+ // something is broken, 3 means retry, the same unlock the partial status
22
+ // names ("/pr-hero review retries it").
23
+
24
+ import type { RunStatus } from "#review/findings";
25
+
26
+ export const EXIT_REVIEW_COMPLETE = 0;
27
+ export const EXIT_REVIEW_FAILED = 1;
28
+ export const EXIT_REVIEW_INCOMPLETE = 3;
29
+
30
+ export type ReviewExitCode =
31
+ | typeof EXIT_REVIEW_COMPLETE
32
+ | typeof EXIT_REVIEW_FAILED
33
+ | typeof EXIT_REVIEW_INCOMPLETE;
34
+
35
+ export function reviewExitCode(input: {
36
+ // Every hunter failed (PipelineResult.sessionFailed).
37
+ readonly sessionFailed: boolean;
38
+ // The pipeline's own statement that no hunter ran on a run meant to review
39
+ // (PipelineResult.noReviewerRan). A re-review that skipped discovery on
40
+ // purpose leaves it false. Absent reads as false.
41
+ readonly noReviewerRan?: boolean;
42
+ // findings.json's `run_status`: "partial" whenever any step failed.
43
+ readonly runStatus: RunStatus;
44
+ // Findings the posting stage could not deliver. 0 when nothing was posted.
45
+ readonly droppedFindingIds: number;
46
+ // Accepted and deliberately NEVER read (spec 2.1): it exists so a test can
47
+ // pass a high blocking count and assert the code does not change, which
48
+ // fails loudly if a future change wires findings into the exit code.
49
+ readonly blockingCount?: number;
50
+ }): ReviewExitCode {
51
+ if (input.sessionFailed || input.noReviewerRan === true) {
52
+ return EXIT_REVIEW_FAILED;
53
+ }
54
+ // WHY a dropped posting outranks an incomplete review: a retry repairs the
55
+ // missing step, but findings that never reached the PR are a delivery
56
+ // failure the author cannot see at all.
57
+ if (input.droppedFindingIds > 0) return EXIT_REVIEW_FAILED;
58
+ if (input.runStatus === "partial") return EXIT_REVIEW_INCOMPLETE;
59
+ return EXIT_REVIEW_COMPLETE;
60
+ }
@@ -661,9 +661,9 @@ export function deriveTier(
661
661
  // gets this power — `inconclusive` still blocks, mirroring the rule that
662
662
  // `refuted` requires positive disproof. Silence is not a demotion.
663
663
  if (finding.refuter_verdict === "downgraded-latent") return "advisory";
664
- // The blocking tier does not change the job's exit code: ciExitCode stays 0
665
- // on findings, and "blocking findings exit 0 in CI — reviewer, not a merge
666
- // gate" (test/ci/review.test.ts) pins that. What it does change is the pr-hero commit status, and only in
664
+ // The blocking tier does not change the job's exit code: reviewExitCode
665
+ // never reads findings (src/review/exit-code.ts, spec 2.1), and
666
+ // test/review/exit-code.test.ts pins that. What it does change is the pr-hero commit status, and only in
667
667
  // gate mode (the CI default, S4): there that status, which is the required
668
668
  // check, goes red on blocking findings. In advisory mode the tier buys only
669
669
  // the report's loudest register: the red badge, the "Blocking" section, and
@@ -613,7 +613,15 @@ questions; use it first.
613
613
  The size gate is a COST gate, not a quality gate: past its limits a diff bills
614
614
  several times more with a much wider spread, so Crysk skips it unattended
615
615
  rather than guess at the bill. An interactive run asks first. It says nothing
616
- about how well a large diff reviews.`;
616
+ about how well a large diff reviews.
617
+
618
+ Exit codes of a review (findings never change them):
619
+ 0 the review is complete, with or without findings
620
+ 1 no usable review: every hunter failed, no reviewer ran, or findings
621
+ that should have been posted were dropped
622
+ 2 usage error
623
+ 3 the review is incomplete: a step failed (a hunter, the refuter,
624
+ verify, adjudicate, or the time ceiling). Run it again.`;
617
625
 
618
626
  const VALUE_FLAGS = new Set([
619
627
  "--repo",
@@ -8,6 +8,7 @@
8
8
  import { EFFORT_LEVELS, type Effort, isEffort } from "#model/catalog";
9
9
  import type { ModelProfile } from "#model/profiles";
10
10
  import { PROFILE_PROMPT_FILES } from "../asset-manifest";
11
+ import { resolveBundledFilePath } from "../assets";
11
12
  import { wrapBlock } from "./boundary";
12
13
 
13
14
  // Fixed order — the lab's promptSetFingerprint hashes the concatenated files
@@ -69,7 +70,11 @@ export function profilePromptSet(
69
70
  `model profile "${profile.id}" names prompt set "${profile.promptSet}", which this build does not have`,
70
71
  );
71
72
  }
72
- for (const [file, filePath] of Object.entries(set)) {
73
+ for (const [file, bundledPath] of Object.entries(set)) {
74
+ // Set paths come from bundled `with { type: "file" }` imports, which an npm
75
+ // build emits as cwd-relative strings; resolve them against the module like
76
+ // every other bundled asset (src/assets.ts).
77
+ const filePath = resolveBundledFilePath(bundledPath);
73
78
  if (defaultAgentFiles.includes(file)) {
74
79
  overrides.agents[file] = filePath;
75
80
  } else if (isEnginePromptFile(file)) {
@@ -26,6 +26,7 @@ import { engineIdentity } from "#git/identity";
26
26
  import { countFreeRouteSteps } from "#model/free-discovery";
27
27
  import { planRoutes } from "#model/routing";
28
28
  import { reviewPr } from "#pr/review-pr";
29
+ import { reviewExitCode } from "#review/exit-code";
29
30
  import type { Telemetry } from "#review/findings";
30
31
  import { type PipelineResult, runPipeline } from "#review/pipeline";
31
32
  import {
@@ -537,6 +538,14 @@ export async function review(options: CliOptions): Promise<number> {
537
538
  })) {
538
539
  log(line);
539
540
  }
540
- if (result.sessionFailed) return 1;
541
- return 0;
541
+ // The same table as PR and CI mode (src/review/exit-code.ts). Local mode
542
+ // posts nothing, so nothing can be dropped.
543
+ return reviewExitCode({
544
+ sessionFailed: result.sessionFailed,
545
+ ...(result.noReviewerRan === undefined
546
+ ? {}
547
+ : { noReviewerRan: result.noReviewerRan }),
548
+ runStatus: doc.run_status,
549
+ droppedFindingIds: 0,
550
+ });
542
551
  }
@@ -4,8 +4,9 @@
4
4
  // `dismissed` reply to a blocking finding, and posts the ruling as a reply on the
5
5
  // finding's thread. That reply is the only CI verdict the gate trusts, and only
6
6
  // when all of these hold:
7
- // - GitHub says the reply's author is CI_ADJUDICATOR_LOGIN. The marker's own
8
- // text proves nothing, since anyone can type it;
7
+ // - GitHub says the reply's author is CI_ADJUDICATOR_LOGIN, or the GitHub App bot
8
+ // that posted the finding itself (trustedAdjudicator). The marker's own text
9
+ // proves nothing, since anyone can type it;
9
10
  // - its `reply=` names the comment id of the thread's LATEST triage reply, so a
10
11
  // newer dismissal is pending again;
11
12
  // - that triage reply is `dismissed`, and the verdict is `upheld` (D5: only a
@@ -21,6 +22,38 @@ export const ADJUDICATION_MARKER_PREFIX = `${MARKER_OPEN}adjudication `;
21
22
  // The login GitHub stamps on comments posted with the workflow's GITHUB_TOKEN.
22
23
  export const CI_ADJUDICATOR_LOGIN = "github-actions[bot]";
23
24
 
25
+ // A GitHub App's bot login. A person's login can never end in `[bot]`.
26
+ const APP_BOT_LOGIN = /^[A-Za-z0-9](?:[A-Za-z0-9-]*[A-Za-z0-9])?\[bot\]$/;
27
+
28
+ // Who a read trusts to rule (odd/tasks/platform-commands.md in the platform repo, S5).
29
+ // - `ci`: a run's own read, which decides the gate. Standalone only github-actions[bot] rules. A
30
+ // run a Crysk platform dispatched posts as the platform App's bot, the login the platform's
31
+ // token exchange returned (CRYSK_PLATFORM_ACTOR), so that login rules too, on a finding it posted.
32
+ // - `local`: `crysk status` on a developer's machine, which never learns the platform's login. It
33
+ // trusts an App's bot ruling on a finding that same bot posted. That read only advises the next
34
+ // action; the gate's own read is `ci`.
35
+ export type AdjudicatorTrust =
36
+ | { kind: "ci"; platformActor?: string }
37
+ | { kind: "local" };
38
+
39
+ export const CI_ADJUDICATOR_TRUST: AdjudicatorTrust = { kind: "ci" };
40
+
41
+ // WHY the finding's own author: another App installed on the repository can post the marker too,
42
+ // but only the bot that posted the finding rules on it. A person can type the marker under a
43
+ // finding they posted, so the author must be an App's bot.
44
+ export function trustedAdjudicator(
45
+ rulingAuthor: string | undefined,
46
+ findingAuthor: string | undefined,
47
+ trust: AdjudicatorTrust,
48
+ ): boolean {
49
+ if (rulingAuthor === undefined) return false;
50
+ if (rulingAuthor === CI_ADJUDICATOR_LOGIN) return true;
51
+ if (rulingAuthor !== findingAuthor || !APP_BOT_LOGIN.test(rulingAuthor)) {
52
+ return false;
53
+ }
54
+ return trust.kind === "local" || rulingAuthor === trust.platformActor;
55
+ }
56
+
24
57
  // The most adjudicator calls one run makes. A pending dismissal past the cap
25
58
  // stays pending, still counted, and the next recheck takes it.
26
59
  export const MAX_ADJUDICATIONS_PER_RUN = 5;
package/src/ui/result.ts CHANGED
@@ -540,7 +540,7 @@ export function renderResult(input: ResultInput): string[] {
540
540
  );
541
541
  }
542
542
  // The line that outranks every count above it. The exit code is the SHELL's
543
- // (return 1 in local mode, postingExitCode in PR mode) — this only says so.
543
+ // (reviewExitCode, src/review/exit-code.ts: 1 here) — this only says so.
544
544
  if (input.sessionFailed) {
545
545
  lines.push(
546
546
  "",
package/src/updater.ts CHANGED
@@ -55,7 +55,7 @@ export function detectInstallMethod(
55
55
  ): InstallMethod {
56
56
  const home = options.home ?? os.homedir();
57
57
  const execPath = options.execPath ?? process.execPath;
58
- const version = options.version ?? "0.6.0";
58
+ const version = options.version ?? "0.6.1";
59
59
  const exists = options.exists ?? existsSync;
60
60
 
61
61
  const isSource =
@@ -11,6 +11,10 @@
11
11
  import path from "node:path";
12
12
  import { isFullCommitId } from "#git/refs";
13
13
  import { PR_COMMENT_MARKER_PREFIX } from "#pr/preflight";
14
+ import {
15
+ EXIT_REVIEW_COMPLETE,
16
+ EXIT_REVIEW_INCOMPLETE,
17
+ } from "#review/exit-code";
14
18
  import type { NumstatFile } from "#review/preflight";
15
19
  import {
16
20
  DEFAULT_SIZE_GATE,
@@ -1096,9 +1100,21 @@ export function launchedLine(
1096
1100
  return `${ts} launched pr=${pr} repo=${repoBase} head=${headSha.slice(0, 8)}`;
1097
1101
  }
1098
1102
 
1103
+ // What the watcher makes of a review child's exit code, by the review's own
1104
+ // table (src/review/exit-code.ts). "incomplete" is exit 3: the review ran and
1105
+ // may have posted findings, but a step failed, so it is neither a clean "ok"
1106
+ // nor a "failed" whose counts mean nothing.
1107
+ export type ReviewResult = "ok" | "incomplete" | "failed";
1108
+
1109
+ export function reviewOutcomeResult(exitCode: number): ReviewResult {
1110
+ if (exitCode === EXIT_REVIEW_COMPLETE) return "ok";
1111
+ if (exitCode === EXIT_REVIEW_INCOMPLETE) return "incomplete";
1112
+ return "failed";
1113
+ }
1114
+
1099
1115
  export interface ReviewOutcome {
1100
1116
  pr: number;
1101
- ok: boolean;
1117
+ result: ReviewResult;
1102
1118
  exitCode: number;
1103
1119
  counts: { blocking: number; advisory: number } | null;
1104
1120
  }
@@ -1108,13 +1124,17 @@ export function outcomeLine(
1108
1124
  repoBase: string,
1109
1125
  outcome: ReviewOutcome,
1110
1126
  ): string {
1111
- const status = outcome.ok ? "status=ok" : "status=failed";
1112
- const detail = outcome.ok
1113
- ? outcome.counts === null
1127
+ const counts =
1128
+ outcome.counts === null
1114
1129
  ? ""
1115
- : ` blocking=${outcome.counts.blocking} advisory=${outcome.counts.advisory}`
1116
- : ` exit=${outcome.exitCode}`;
1117
- return `${ts} outcome pr=${outcome.pr} repo=${repoBase} ${status}${detail}`;
1130
+ : ` blocking=${outcome.counts.blocking} advisory=${outcome.counts.advisory}`;
1131
+ const detail =
1132
+ outcome.result === "ok"
1133
+ ? counts
1134
+ : outcome.result === "incomplete"
1135
+ ? ` exit=${outcome.exitCode}${counts}`
1136
+ : ` exit=${outcome.exitCode}`;
1137
+ return `${ts} outcome pr=${outcome.pr} repo=${repoBase} status=${outcome.result}${detail}`;
1118
1138
  }
1119
1139
 
1120
1140
  // Counts today's launches back out of the raw log text. dayPrefix is the
@@ -1267,9 +1287,15 @@ export function osascriptNotifyArgs(title: string, text: string): string[] {
1267
1287
  }
1268
1288
 
1269
1289
  export function outcomeNotificationText(outcome: ReviewOutcome): string {
1270
- if (!outcome.ok) {
1290
+ if (outcome.result === "failed") {
1271
1291
  return `PR #${outcome.pr} review failed (exit ${outcome.exitCode})`;
1272
1292
  }
1293
+ if (outcome.result === "incomplete") {
1294
+ const prefix = `PR #${outcome.pr} review incomplete, run it again`;
1295
+ return outcome.counts === null
1296
+ ? `${prefix} (counts unavailable)`
1297
+ : `${prefix}: ${outcome.counts.blocking} blocking, ${outcome.counts.advisory} advisory`;
1298
+ }
1273
1299
  if (outcome.counts === null) {
1274
1300
  return `PR #${outcome.pr} reviewed (counts unavailable)`;
1275
1301
  }
@@ -102,6 +102,7 @@ import {
102
102
  removeWatchRepo,
103
103
  renderWatchPlist,
104
104
  renderWatchStatus,
105
+ reviewOutcomeResult,
105
106
  skipLine,
106
107
  type TickDecision,
107
108
  type TickLaunch,
@@ -860,14 +861,17 @@ async function runTick(
860
861
  );
861
862
  const exitCode = await proc.exited;
862
863
 
864
+ // An incomplete review (exit 3) still ran and may have posted findings, so
865
+ // its counts are read like a complete one's; only a failed run has none.
866
+ const result = reviewOutcomeResult(exitCode);
863
867
  const outcome: ReviewOutcome = {
864
868
  pr: launch.pr,
865
- ok: exitCode === 0,
869
+ result,
866
870
  exitCode,
867
871
  counts:
868
- exitCode === 0
869
- ? await readTierCounts(launched?.runsRoot, launch.pr, launch.head)
870
- : null,
872
+ result === "failed"
873
+ ? null
874
+ : await readTierCounts(launched?.runsRoot, launch.pr, launch.head),
871
875
  };
872
876
  await appendLog(
873
877
  paths.logPath,