crysk 0.6.1 → 0.8.0
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 +28 -22
- package/dist/SKILL-1q3658n0.md +74 -0
- package/dist/SKILL-25j4q9b8.md +143 -0
- package/dist/SKILL-32cmme3j.md +143 -0
- package/dist/SKILL-4747bhne.md +143 -0
- package/dist/SKILL-4jc3sv23.md +74 -0
- package/dist/SKILL-c0492rch.md +540 -0
- package/dist/SKILL-d1nzkhhr.md +74 -0
- package/dist/SKILL-f8p7kvx0.md +143 -0
- package/dist/SKILL-gxd45g38.md +553 -0
- package/dist/SKILL-rec41991.md +66 -0
- package/dist/SKILL-zze8zj5g.md +536 -0
- package/dist/adjudicator-eckqh14e.md +78 -0
- package/dist/ci-admission-qn797a6c.md +128 -0
- package/dist/cli.js +18608 -18523
- package/dist/consolidate-4xw1pkg7.md +24 -0
- package/dist/deep-review-logic-dgr112y1.md +196 -0
- package/dist/deep-review-logic-fck5z4xe.md +197 -0
- package/dist/deep-review-logic-h1fpamfh.md +192 -0
- package/dist/deep-review-logic-m4vk5683.md +196 -0
- package/dist/deep-review-logic-z63kmgmc.md +157 -0
- package/dist/deep-review-parity-7x7s5d81.md +148 -0
- package/dist/deep-review-parity-9wxmynsc.md +115 -0
- package/dist/deep-review-parity-bwwe7cx6.md +152 -0
- package/dist/deep-review-parity-e50pw3pk.md +117 -0
- package/dist/deep-review-parity-ennmm9m1.md +151 -0
- package/{prompts/default/deep-review-parity.md → dist/deep-review-parity-h7y6y2he.md} +14 -16
- package/dist/deep-review-parity-xg6wy69c.md +151 -0
- package/dist/deep-review-reliability-6h3c4gmw.md +125 -0
- package/dist/deep-review-reliability-dcjab69x.md +122 -0
- package/dist/deep-review-reliability-ec7b5pbk.md +159 -0
- package/dist/deep-review-reliability-hvasvmt9.md +153 -0
- package/dist/deep-review-reliability-mjnayxmt.md +158 -0
- package/dist/deep-review-reliability-mm7fnayf.md +124 -0
- package/dist/deep-review-reliability-xk4xvc2w.md +158 -0
- package/dist/deep-review-resilience-0esd5qeh.md +148 -0
- package/dist/deep-review-resilience-envgcmpy.md +115 -0
- package/dist/deep-review-resilience-fgx8ythf.md +151 -0
- package/dist/deep-review-resilience-mq5qa6ad.md +152 -0
- package/dist/deep-review-resilience-pwh1gawp.md +151 -0
- package/dist/deep-review-resilience-s9n21dkx.md +117 -0
- package/dist/deep-review-resilience-vz4qskea.md +119 -0
- package/dist/deep-review-security-1vwjdc9q.md +205 -0
- package/dist/deep-review-security-x1cc1z00.md +205 -0
- package/dist/index.js +590 -485
- package/dist/opencode-ci-jc8mf4wn.md +106 -0
- package/dist/opencode-ci-qrcv7a99.md +106 -0
- package/dist/opencode-ci-rdpcc7xc.md +106 -0
- package/dist/opencode-ci-yt0k74t0.md +106 -0
- package/dist/review-refuter-7qffwtzz.md +77 -0
- package/dist/verifier-tyh8r45v.md +40 -0
- package/dist/verifier-xf4xj701.md +40 -0
- package/dist/workflow-andv2dvm.yml +158 -0
- package/package.json +1 -1
- package/prompts/consolidate.md +1 -1
- package/prompts/default/PROVENANCE.md +6 -0
- package/prompts/default/deep-review-logic.md +56 -15
- package/prompts/default/deep-review-reliability.md +64 -30
- package/prompts/default/deep-review-resilience.md +71 -37
- package/prompts/default/deep-review-security.md +205 -0
- package/prompts/default/review-refuter.md +1 -1
- package/prompts/verifier.md +40 -0
- package/skills/crysk/SKILL.md +6 -6
- package/skills/crysk-ci-setup/SKILL.md +32 -33
- package/skills/crysk-ci-setup/assets/workflow.yml +5 -10
- package/skills/crysk-ci-setup/references/ci-admission.md +5 -5
- package/skills/crysk-ci-setup/references/opencode-ci.md +6 -6
- package/skills/crysk-loop/SKILL.md +37 -54
- package/skills/crysk-loop/adjudicator.md +10 -18
- package/skills/crysk-pr/SKILL.md +1 -1
- package/src/asset-manifest.ts +6 -6
- package/src/assets.ts +4 -4
- package/src/brand.ts +20 -30
- package/src/ci/gates.ts +155 -192
- package/src/ci/outcome.ts +4 -31
- package/src/ci/platform-limits.ts +135 -0
- package/src/ci/platform-report.ts +212 -126
- package/src/ci/platform.ts +28 -26
- package/src/ci/reporter.ts +25 -26
- package/src/ci/review-admission.ts +1 -2
- package/src/ci/review-levels.ts +12 -53
- package/src/ci/setup.ts +96 -66
- package/src/ci/trace-bundle.ts +11 -27
- package/src/ci/trace-transcripts.ts +4 -6
- package/src/ci/trigger.ts +3 -5
- package/src/cli.ts +6 -5
- package/src/commands/ci-setup-questions.ts +2 -55
- package/src/commands/ci-setup.ts +2 -26
- package/src/commands/triage.ts +131 -38
- package/src/commands/usage.ts +0 -1
- package/src/compare/compare.ts +3 -3
- package/src/compare/floor-test.ts +30 -201
- package/src/compare/hunter-exam.ts +631 -0
- package/src/config/config.ts +6 -30
- package/src/diversity/accounting.ts +0 -26
- package/src/diversity/benchmark.ts +1 -17
- package/src/diversity/identity.ts +0 -14
- package/src/diversity/pipeline-integration.ts +0 -29
- package/src/doctor.ts +3 -6
- package/src/execution/claude-session-trace.ts +7 -2
- package/src/execution/contracts.ts +15 -9
- package/src/execution/harness.ts +165 -187
- package/src/execution/spend-limiter.ts +39 -131
- package/src/execution/usage-facts.ts +397 -0
- package/src/execution/usage-normalized.ts +33 -86
- package/src/harness/materiality-judge.ts +0 -5
- package/src/harness/pipeline-plan.ts +0 -3
- package/src/harness/replay-rows.ts +0 -5
- package/src/harness/route-runner.ts +2 -2
- package/src/harness/single-step.ts +3 -9
- package/src/index.ts +1 -1
- package/src/mcp/preflight.ts +2 -2
- package/src/model/catalog.ts +1 -3
- package/src/model/free-discovery.ts +11 -78
- package/src/model/model-name.ts +16 -5
- package/src/model/profiles/claude.ts +1 -2
- package/src/model/profiles/openai.ts +1 -2
- package/src/model/profiles.ts +2 -3
- package/src/model/provider-capabilities.ts +16 -55
- package/src/model/routing.ts +39 -36
- package/src/pr/admission.ts +6 -6
- package/src/pr/answered-findings.ts +8 -4
- package/src/pr/ci-publish.ts +0 -2
- package/src/pr/discovery.ts +4 -25
- package/src/pr/plan.ts +40 -121
- package/src/pr/pr-status.ts +14 -2
- package/src/pr/pr.ts +1 -1
- package/src/pr/preflight.ts +44 -15
- package/src/pr/publish-outcome.ts +1 -19
- package/src/pr/review-pr.ts +15 -47
- package/src/pr/target.ts +76 -82
- package/src/production-runtime.ts +52 -114
- package/src/progress.ts +0 -54
- package/src/rereview/prepare.ts +4 -1
- package/src/rereview/state.ts +1 -1
- package/src/rereview/verify.ts +15 -9
- package/src/review/boundary.ts +5 -8
- package/src/review/drafts.ts +3 -3
- package/src/review/findings-conformance.ts +1 -1
- package/src/review/findings.ts +0 -4
- package/src/review/pipeline.ts +194 -407
- package/src/review/preflight.ts +175 -201
- package/src/review/prompt-set.ts +1 -2
- package/src/review/report.ts +2 -162
- package/src/review/review.ts +25 -74
- package/src/review/route-preflight.ts +19 -35
- package/src/review/run.ts +37 -73
- package/src/review/size-gate.ts +3 -3
- package/src/review/spec.ts +6 -11
- package/src/review/step-runner.ts +4 -4
- package/src/runner-authority.ts +10 -9
- package/src/security/credential-broker.ts +5 -6
- package/src/server/preflight.ts +0 -1
- package/src/server/server.ts +17 -6
- package/src/store/activity.ts +1 -2
- package/src/store/backfill-preflight.ts +0 -3
- package/src/store/backfill.ts +0 -1
- package/src/store/metrics-preflight.ts +4 -14
- package/src/store/metrics.ts +5 -5
- package/src/store/preflight.ts +0 -13
- package/src/store/store.ts +11 -12
- package/src/transport-registry.ts +6 -30
- package/src/transports/claude-code-cli.ts +149 -215
- package/src/transports/opencode-client.ts +4 -24
- package/src/transports/opencode-sdk.ts +17 -112
- package/src/triage/plain-reply.ts +37 -0
- package/src/ui/activity.ts +1 -6
- package/src/ui/config-edit.ts +14 -67
- package/src/ui/config.ts +3 -8
- package/src/ui/plan.ts +11 -198
- package/src/ui/progress.ts +3 -49
- package/src/ui/result.ts +18 -77
- package/src/ui/review-menu.ts +3 -12
- package/src/ui/select.ts +9 -79
- package/src/updater.ts +1 -1
- package/src/usage.ts +1 -7
- package/prompts/scout.md +0 -140
- package/src/ci/platform-estimate.ts +0 -111
- package/src/review/scout.ts +0 -323
package/src/execution/harness.ts
CHANGED
|
@@ -1,5 +1,7 @@
|
|
|
1
|
+
import { createReadStream, existsSync } from "node:fs";
|
|
1
2
|
import { mkdir } from "node:fs/promises";
|
|
2
3
|
import path from "node:path";
|
|
4
|
+
import { createInterface } from "node:readline";
|
|
3
5
|
import { releaseExecutionSnapshot } from "#model/exec-snapshots";
|
|
4
6
|
import {
|
|
5
7
|
type ExecutableAllowlistEntry,
|
|
@@ -53,6 +55,7 @@ import {
|
|
|
53
55
|
type AuthEvent,
|
|
54
56
|
isTransportFailureCause,
|
|
55
57
|
type ProviderTransport,
|
|
58
|
+
type ResolvedModelRoute,
|
|
56
59
|
type StepAdmissionGate,
|
|
57
60
|
type TransportFailureCause,
|
|
58
61
|
type TransportOutcome,
|
|
@@ -81,7 +84,6 @@ import {
|
|
|
81
84
|
beginStep,
|
|
82
85
|
nextCycle,
|
|
83
86
|
type ReserveToken,
|
|
84
|
-
type SettlementDecision,
|
|
85
87
|
type SpendLedger,
|
|
86
88
|
type SpendReservation,
|
|
87
89
|
SpendReservationFencedError,
|
|
@@ -92,12 +94,20 @@ import {
|
|
|
92
94
|
FORMAT_RETRY_REMINDER,
|
|
93
95
|
settlementReceiptPath,
|
|
94
96
|
} from "./step-artifacts";
|
|
97
|
+
import {
|
|
98
|
+
type AttemptUsageFacts,
|
|
99
|
+
claudeAttemptUsageFacts,
|
|
100
|
+
credentialOfBillingMode,
|
|
101
|
+
openCodeAttemptUsageFacts,
|
|
102
|
+
usageFactsPath,
|
|
103
|
+
} from "./usage-facts";
|
|
95
104
|
import {
|
|
96
105
|
envBillsMetered,
|
|
97
106
|
type NormalizedUsage,
|
|
98
107
|
normalizeUnavailableUsage,
|
|
99
108
|
projectLegacyUsage,
|
|
100
109
|
sumNormalizedUsage,
|
|
110
|
+
type UsageBillingMode,
|
|
101
111
|
} from "./usage-normalized";
|
|
102
112
|
|
|
103
113
|
// §7 line 416: injected sleep for the `rate_limit` capped-exponential
|
|
@@ -249,13 +259,6 @@ export interface StepExecutionHarnessOptions {
|
|
|
249
259
|
// every existing caller (PR5a's own tests included) keeps its ledger-free
|
|
250
260
|
// shape unmodified.
|
|
251
261
|
readonly spendLedger?: SpendLedger;
|
|
252
|
-
// What each attempt reserves BEFORE its actual cost is known. Estimating a
|
|
253
|
-
// real per-step dollar figure from prompt/model/provider pricing is a
|
|
254
|
-
// caller concern this slice does not attempt (same framing as PR5a's own
|
|
255
|
-
// `rateLimitBucketId` default: deriving the real number belongs to
|
|
256
|
-
// whoever composes the run, not to this harness). Defaults to 0 so an
|
|
257
|
-
// unconfigured caller reserves nothing rather than an invented ceiling.
|
|
258
|
-
readonly reservedUsdPerAttempt?: number;
|
|
259
262
|
// #279: the BINDING's own reserve decision — true for a metered or
|
|
260
263
|
// free-credential route (production-runtime.ts's `reservesSpend`),
|
|
261
264
|
// independent of whether `spendLedger` is configured at all. Defaults to
|
|
@@ -328,7 +331,7 @@ const ENV_PASSTHROUGH = [
|
|
|
328
331
|
// disagreeing about the same variable. pr-hero's first real CI self-review
|
|
329
332
|
// (2026-08-27) passed the gate, passed the size gate, then lost all three
|
|
330
333
|
// hunters and the summarizer in three seconds to "Not logged in · Please run
|
|
331
|
-
// /login",
|
|
334
|
+
// /login", because this projection dropped the token the action
|
|
332
335
|
// had just handed it. It is the same credential class as ANTHROPIC_API_KEY
|
|
333
336
|
// directly above, and it is the ONLY route on Linux: credential projection
|
|
334
337
|
// needs darwin + /usr/bin/security, so every CI runner falls back to exactly
|
|
@@ -347,9 +350,9 @@ const ENV_PASSTHROUGH = [
|
|
|
347
350
|
// untouched on exactly the path whose purpose is isolating the credential.
|
|
348
351
|
//
|
|
349
352
|
// The leak was inert while nothing read it back. #177 ended that: the Claude
|
|
350
|
-
// CLI transport now derives `billingMode
|
|
351
|
-
// (`
|
|
352
|
-
// projected subscription credential got the review filed as metered
|
|
353
|
+
// CLI transport now derives `billingMode` from this very env
|
|
354
|
+
// (`claudeCliBillingMode` → `envBillsMetered`), so a developer running on the
|
|
355
|
+
// projected subscription credential got the review filed as metered on
|
|
353
356
|
// the strength of a key the projection was supposed to have replaced. Both
|
|
354
357
|
// readings of the leak are defects — either the CLI honours the leaked key
|
|
355
358
|
// and the isolation promise is false, or it prefers the projected OAuth and
|
|
@@ -413,7 +416,7 @@ function notifyRetry(step: StepSpec, info: RetryInfo): void {
|
|
|
413
416
|
const DESTROY_FAILED_ANNOTATION =
|
|
414
417
|
"[pr-hero] credential projection destroy failed";
|
|
415
418
|
const AMBIENT_METERED_ANNOTATION =
|
|
416
|
-
"[pr-hero] ambient credential bills metered — this attempt reserves against the spend ledger and fences its bucket if
|
|
419
|
+
"[pr-hero] ambient credential bills metered — this attempt reserves against the spend ledger and fences its bucket if its usage cannot be confirmed";
|
|
417
420
|
// The class is variable in the shape even though only
|
|
418
421
|
// missing_subscription_record reaches the append today: the shape is what
|
|
419
422
|
// the harness writes, the class list is a policy that may grow.
|
|
@@ -438,20 +441,96 @@ function copyClaudeSessionTrace(
|
|
|
438
441
|
).catch((): ClaudeTraceMiss => "session trace write failed");
|
|
439
442
|
}
|
|
440
443
|
|
|
441
|
-
|
|
442
|
-
step: StepSpec,
|
|
443
|
-
request: TransportRequest,
|
|
444
|
-
attempt: number,
|
|
444
|
+
function withTraceMissAnnotation(
|
|
445
445
|
outcome: TransportOutcome,
|
|
446
|
-
|
|
447
|
-
|
|
448
|
-
if (
|
|
446
|
+
miss: ClaudeTraceMiss | undefined,
|
|
447
|
+
): TransportOutcome {
|
|
448
|
+
if (miss === undefined) return outcome;
|
|
449
449
|
return {
|
|
450
450
|
...outcome,
|
|
451
|
-
stderrTail: `${outcome.stderrTail}\n${claudeTraceCaptureAnnotation(
|
|
451
|
+
stderrTail: `${outcome.stderrTail}\n${claudeTraceCaptureAnnotation(miss)}`,
|
|
452
452
|
};
|
|
453
453
|
}
|
|
454
454
|
|
|
455
|
+
// platform-owned-cost T6: what one attempt measured per model request, for the
|
|
456
|
+
// platform's `usage_facts` (usage-facts.ts). Read here because the session copy
|
|
457
|
+
// and the CLI's per-model summary are only certain to exist now. Like the trace
|
|
458
|
+
// copy it is never the attempt's failure: a facts file that could not be built
|
|
459
|
+
// or written is absent, and the run's report names no requests for the attempt.
|
|
460
|
+
async function writeAttemptUsageFacts(input: {
|
|
461
|
+
step: StepSpec;
|
|
462
|
+
// The route the attempt was sent, for a provider and model the source does not name.
|
|
463
|
+
route: ResolvedModelRoute;
|
|
464
|
+
attempt: number;
|
|
465
|
+
transport: ProviderTransport;
|
|
466
|
+
// The trace copy's answer for this attempt (copyClaudeSessionTrace).
|
|
467
|
+
miss: ClaudeTraceMiss | undefined;
|
|
468
|
+
// Absent on a cancelled attempt, which has no outcome.
|
|
469
|
+
outcome: TransportOutcome | undefined;
|
|
470
|
+
billingMode: UsageBillingMode;
|
|
471
|
+
}): Promise<void> {
|
|
472
|
+
const { step, route, attempt, outcome } = input;
|
|
473
|
+
try {
|
|
474
|
+
const credential = credentialOfBillingMode(input.billingMode);
|
|
475
|
+
if (credential === undefined) return;
|
|
476
|
+
let facts: AttemptUsageFacts | undefined;
|
|
477
|
+
if (input.transport.backend === "claude-code") {
|
|
478
|
+
const trace = claudeSessionTracePath(step.outPath, step.name, attempt);
|
|
479
|
+
const copied = input.miss === undefined && existsSync(trace);
|
|
480
|
+
// No copy and no miss is an attempt that spawned no CLI: nothing ran.
|
|
481
|
+
if (
|
|
482
|
+
!copied &&
|
|
483
|
+
input.miss === undefined &&
|
|
484
|
+
outcome?.modelUsage === undefined
|
|
485
|
+
) {
|
|
486
|
+
return;
|
|
487
|
+
}
|
|
488
|
+
const claude = (session: AsyncIterable<string> | undefined) =>
|
|
489
|
+
claudeAttemptUsageFacts({
|
|
490
|
+
attempt,
|
|
491
|
+
credential,
|
|
492
|
+
provider: route.provider,
|
|
493
|
+
session,
|
|
494
|
+
summary: outcome?.modelUsage,
|
|
495
|
+
});
|
|
496
|
+
try {
|
|
497
|
+
// A line at a time: the copy can be 64 MiB (see claude-session-trace.ts).
|
|
498
|
+
facts = await claude(
|
|
499
|
+
copied
|
|
500
|
+
? createInterface({
|
|
501
|
+
input: createReadStream(trace, "utf8"),
|
|
502
|
+
crlfDelay: Number.POSITIVE_INFINITY,
|
|
503
|
+
})
|
|
504
|
+
: undefined,
|
|
505
|
+
);
|
|
506
|
+
} catch {
|
|
507
|
+
// The copy could not be read: report it as missing (a `no_session` gap), so
|
|
508
|
+
// the CLI's per-model summary still stands in for its requests.
|
|
509
|
+
facts = await claude(undefined);
|
|
510
|
+
}
|
|
511
|
+
} else if (
|
|
512
|
+
input.transport.backend === "opencode" &&
|
|
513
|
+
outcome !== undefined
|
|
514
|
+
) {
|
|
515
|
+
facts = openCodeAttemptUsageFacts({
|
|
516
|
+
attempt,
|
|
517
|
+
credential,
|
|
518
|
+
route: { provider: route.provider, model: route.modelSnapshot },
|
|
519
|
+
capture: outcome.diagnosticEvidence,
|
|
520
|
+
totals: outcome.usage.tokens,
|
|
521
|
+
});
|
|
522
|
+
}
|
|
523
|
+
if (facts === undefined) return;
|
|
524
|
+
await writeJsonAtomically(
|
|
525
|
+
usageFactsPath(step.outPath, step.name, attempt),
|
|
526
|
+
facts,
|
|
527
|
+
{ compact: true },
|
|
528
|
+
);
|
|
529
|
+
} catch {
|
|
530
|
+
// See above: the facts are absent, the attempt is unchanged.
|
|
531
|
+
}
|
|
532
|
+
}
|
|
533
|
+
|
|
455
534
|
export function isHarnessStderrAnnotation(line: string): boolean {
|
|
456
535
|
const trimmed = line.trim();
|
|
457
536
|
return (
|
|
@@ -588,7 +667,6 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
588
667
|
private readonly attemptAdmissionGate?: AttemptAdmissionGate;
|
|
589
668
|
private readonly rateLimitBucketId: string;
|
|
590
669
|
private readonly spendLedger?: SpendLedger;
|
|
591
|
-
private readonly reservedUsdPerAttempt: number;
|
|
592
670
|
private readonly reservesSpend: boolean;
|
|
593
671
|
private readonly degradedProjectionBucketId?: string;
|
|
594
672
|
private readonly attemptAdmissionTimeoutMs: number;
|
|
@@ -621,7 +699,6 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
621
699
|
this.rateLimitBucketId =
|
|
622
700
|
options.rateLimitBucketId ?? DEFAULT_RATE_LIMIT_BUCKET_ID;
|
|
623
701
|
this.spendLedger = options.spendLedger;
|
|
624
|
-
this.reservedUsdPerAttempt = options.reservedUsdPerAttempt ?? 0;
|
|
625
702
|
// #279: default `true` — see the WHY on `StepExecutionHarnessOptions.
|
|
626
703
|
// reservesSpend` for why an unset flag must keep every pre-existing
|
|
627
704
|
// `spendLedger`-configuring caller reserving unconditionally.
|
|
@@ -1162,7 +1239,7 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
1162
1239
|
// - attempt vs attempt: every retry receives its own receipt, so a later
|
|
1163
1240
|
// attempt never clobbers an earlier one's rejected-event / fence record.
|
|
1164
1241
|
// - step vs step: every step of a run shares ONE steps/ directory
|
|
1165
|
-
// (hunters, summarizer, refuters
|
|
1242
|
+
// (hunters, summarizer, refuters all write beside each other),
|
|
1166
1243
|
// so a name keyed only on the attempt made each step overwrite the
|
|
1167
1244
|
// previous step's receipt. That is not hypothetical — a four-step run
|
|
1168
1245
|
// left four attempt logs and exactly one receipt on disk, and three
|
|
@@ -1430,14 +1507,7 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
1430
1507
|
finalText: "",
|
|
1431
1508
|
usage: normalizeUnavailableUsage({
|
|
1432
1509
|
wallMs: harnessWatchdogMs ?? 0,
|
|
1433
|
-
|
|
1434
|
-
// answer reads the engine's startup env and could name a credential the run stripped.
|
|
1435
|
-
billingMode:
|
|
1436
|
-
transport.backend === "claude-code"
|
|
1437
|
-
? envBillsMetered(this.projectedEnv)
|
|
1438
|
-
? "metered"
|
|
1439
|
-
: "subscription"
|
|
1440
|
-
: transport.billingMode,
|
|
1510
|
+
billingMode: this.attemptBillingMode(transport),
|
|
1441
1511
|
}),
|
|
1442
1512
|
stderrTail: `Step timed out after ${harnessWatchdogMs}ms`,
|
|
1443
1513
|
timedOut: true,
|
|
@@ -1677,7 +1747,7 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
1677
1747
|
// `projectLegacyUsage` at every return point below, never accumulated
|
|
1678
1748
|
// independently — that is what keeps the two shapes unable to drift.
|
|
1679
1749
|
// `undefined` until the first attempt settles: seeding it with a
|
|
1680
|
-
// billingMode
|
|
1750
|
+
// billingMode-less zero would force `sumNormalizedUsage` to
|
|
1681
1751
|
// collapse the run's real billing mode to "unknown" on the very first sum.
|
|
1682
1752
|
let totalUsageV2: NormalizedUsage | undefined;
|
|
1683
1753
|
let attempts = 0;
|
|
@@ -1971,6 +2041,17 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
1971
2041
|
};
|
|
1972
2042
|
}
|
|
1973
2043
|
|
|
2044
|
+
// The billing mode of an attempt that has no usage record of its own (a watchdog
|
|
2045
|
+
// timeout, a cancellation). A Claude child bills on the credential its env holds,
|
|
2046
|
+
// which the run chose; the transport's own answer reads the engine's startup env
|
|
2047
|
+
// and could name a credential the run stripped.
|
|
2048
|
+
private attemptBillingMode(transport: ProviderTransport): UsageBillingMode {
|
|
2049
|
+
if (transport.backend !== "claude-code") {
|
|
2050
|
+
return transport.billingMode ?? "unknown";
|
|
2051
|
+
}
|
|
2052
|
+
return envBillsMetered(this.projectedEnv) ? "metered" : "subscription";
|
|
2053
|
+
}
|
|
2054
|
+
|
|
1974
2055
|
// D1-08 PR0: the ONE place a single attempt (transient OR format-retry —
|
|
1975
2056
|
// `kind` only affects logging/request framing) executes, gets logged, and
|
|
1976
2057
|
// has its failure resolved to a §7 cause. Replaces the two
|
|
@@ -2049,8 +2130,8 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
2049
2130
|
// a METERED binding of this SAME provider+credential would compute,
|
|
2050
2131
|
// production-runtime.ts) — never this binding's own subscription
|
|
2051
2132
|
// bucket. Fencing the subscription bucket would block the NEXT
|
|
2052
|
-
// non-degraded attempt on the SAME OAuth credential over
|
|
2053
|
-
// credential never
|
|
2133
|
+
// non-degraded attempt on the SAME OAuth credential over usage that
|
|
2134
|
+
// credential never made. NO fallback to `rateLimitBucketId` on this
|
|
2054
2135
|
// branch: the constructor refuses to build a `reservesSpend: false` +
|
|
2055
2136
|
// `spendLedger` harness without `degradedProjectionBucketId` (see the
|
|
2056
2137
|
// throw in the constructor), so a silent wrong-bucket fence can never
|
|
@@ -2067,7 +2148,6 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
2067
2148
|
reservation = await this.spendLedger.reserve(
|
|
2068
2149
|
{
|
|
2069
2150
|
bucketId,
|
|
2070
|
-
reservedUsd: this.reservedUsdPerAttempt,
|
|
2071
2151
|
sessionId: request.sessionId,
|
|
2072
2152
|
attempt,
|
|
2073
2153
|
},
|
|
@@ -2103,86 +2183,16 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
2103
2183
|
harnessWatchdogMs,
|
|
2104
2184
|
});
|
|
2105
2185
|
|
|
2106
|
-
// Step 4: settle
|
|
2107
|
-
//
|
|
2108
|
-
//
|
|
2186
|
+
// Step 4: settle exactly once (or route to markUnresolvedRemote when
|
|
2187
|
+
// usage completeness never became "complete" — coupling 1,
|
|
2188
|
+
// spend-limiter.ts).
|
|
2109
2189
|
if (reservation === undefined) return result;
|
|
2110
2190
|
const finalized = await this.finalizeReservation(
|
|
2111
2191
|
reservation,
|
|
2112
2192
|
result,
|
|
2113
2193
|
transport,
|
|
2114
2194
|
);
|
|
2115
|
-
|
|
2116
|
-
// reason `free_nonzero_cost` means the model flipped from free to
|
|
2117
|
-
// metered mid-flight and the provider priced the work. Unresolved alone
|
|
2118
|
-
// only fences the bucket while the retry loop keeps spending attempts on
|
|
2119
|
-
// the flipped model, so the attempt FAILS CLOSED here with no retry —
|
|
2120
|
-
// `legacy_terminal` breaks the run loop before `decideRetryDisposition`
|
|
2121
|
-
// can resurrect it, and the fence from `markUnresolvedRemote` (already
|
|
2122
|
-
// applied inside `finalizeReservation` — the spend may be real) blocks
|
|
2123
|
-
// the next reserve. Free usage with NO cash figure carries no reason and
|
|
2124
|
-
// releases without fencing (finalizeReservation's free-route fence
|
|
2125
|
-
// scope) — no cash figure means no evidence of billing. Incomplete
|
|
2126
|
-
// usage with a non-free mode still fences: no usage mode to exonerate
|
|
2127
|
-
// it, so fail closed.
|
|
2128
|
-
if (
|
|
2129
|
-
finalized.decision.kind === "unresolved" &&
|
|
2130
|
-
finalized.decision.reason === "free_nonzero_cost"
|
|
2131
|
-
) {
|
|
2132
|
-
const known =
|
|
2133
|
-
finalized.decision.knownUsd ?? finalized.reservation.knownUsd;
|
|
2134
|
-
const flipNote =
|
|
2135
|
-
`[pr-hero] free-route cost flip: provider reported $${known} cash ` +
|
|
2136
|
-
`on a free-declared route; the model may have flipped from free to ` +
|
|
2137
|
-
`metered — re-probe before retrying (no retry attempted)`;
|
|
2138
|
-
if (result.kind === "cancelled") {
|
|
2139
|
-
return {
|
|
2140
|
-
kind: "failed",
|
|
2141
|
-
outcome: {
|
|
2142
|
-
completion: "failed",
|
|
2143
|
-
protocolIntegrity: "unverified",
|
|
2144
|
-
finalText: "",
|
|
2145
|
-
usage: normalizeUnavailableUsage({ wallMs: 0 }),
|
|
2146
|
-
stderrTail: flipNote,
|
|
2147
|
-
},
|
|
2148
|
-
resolution: { kind: "legacy_terminal" },
|
|
2149
|
-
reservation: finalized.reservation,
|
|
2150
|
-
};
|
|
2151
|
-
}
|
|
2152
|
-
// A DELIVERED attempt already persisted its parsed output
|
|
2153
|
-
// (writeJsonAtomically in runAdmittedAttempt) and the loop returns ok
|
|
2154
|
-
// on delivered without consulting the resolution — so rewriting it to
|
|
2155
|
-
// failed would claim an artifact was never written. Preserve the
|
|
2156
|
-
// delivered kind and append the flip note to stderrTail instead. The
|
|
2157
|
-
// fence is already applied above, and no retry follows delivered
|
|
2158
|
-
// anyway, so nothing about fail-closed is weakened: the next reserve
|
|
2159
|
-
// on this bucket still refuses.
|
|
2160
|
-
if (result.kind === "delivered") {
|
|
2161
|
-
return {
|
|
2162
|
-
kind: "delivered",
|
|
2163
|
-
outcome: {
|
|
2164
|
-
...result.outcome,
|
|
2165
|
-
stderrTail: [result.outcome.stderrTail, flipNote]
|
|
2166
|
-
.filter(Boolean)
|
|
2167
|
-
.join("\n"),
|
|
2168
|
-
},
|
|
2169
|
-
parsed: result.parsed,
|
|
2170
|
-
reservation: finalized.reservation,
|
|
2171
|
-
};
|
|
2172
|
-
}
|
|
2173
|
-
return {
|
|
2174
|
-
kind: "failed",
|
|
2175
|
-
outcome: {
|
|
2176
|
-
...result.outcome,
|
|
2177
|
-
stderrTail: [result.outcome.stderrTail, flipNote]
|
|
2178
|
-
.filter(Boolean)
|
|
2179
|
-
.join("\n"),
|
|
2180
|
-
},
|
|
2181
|
-
resolution: { kind: "legacy_terminal" },
|
|
2182
|
-
reservation: finalized.reservation,
|
|
2183
|
-
};
|
|
2184
|
-
}
|
|
2185
|
-
return { ...result, reservation: finalized.reservation };
|
|
2195
|
+
return { ...result, reservation: finalized };
|
|
2186
2196
|
} catch (error) {
|
|
2187
2197
|
if (
|
|
2188
2198
|
error instanceof ConcurrencyAdmissionAbortedError ||
|
|
@@ -2214,111 +2224,64 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
2214
2224
|
// `settlementFromUsage` and applies whichever CAS transition it names.
|
|
2215
2225
|
// `result.kind === "cancelled"` here means the provider attempt WAS
|
|
2216
2226
|
// invoked (executeSession calls the transport unconditionally) but no
|
|
2217
|
-
// usage ever arrived — that case can never settle
|
|
2218
|
-
//
|
|
2219
|
-
//
|
|
2220
|
-
// Returns the decision alongside the terminal reservation so `runAttempt`
|
|
2221
|
-
// can fail closed on `free_nonzero_cost` (#182 follow-up) — the only caller.
|
|
2227
|
+
// usage ever arrived — that case can never settle either, so it takes the
|
|
2228
|
+
// same unresolved path.
|
|
2222
2229
|
private async finalizeReservation(
|
|
2223
2230
|
reservation: SpendReservation,
|
|
2224
2231
|
result: AttemptRunResult,
|
|
2225
2232
|
transport?: ProviderTransport,
|
|
2226
|
-
): Promise<{
|
|
2227
|
-
readonly reservation: SpendReservation;
|
|
2228
|
-
readonly decision: SettlementDecision;
|
|
2229
|
-
}> {
|
|
2233
|
+
): Promise<SpendReservation> {
|
|
2230
2234
|
const ledger = this.spendLedger;
|
|
2231
2235
|
// Unreachable via `runAttempt` (it returns before calling here when no
|
|
2232
2236
|
// reservation was opened, which is exactly when no ledger is configured),
|
|
2233
|
-
// but kept total:
|
|
2234
|
-
|
|
2235
|
-
if (ledger === undefined) {
|
|
2236
|
-
const usage =
|
|
2237
|
-
result.kind === "cancelled" ? undefined : result.outcome.usage;
|
|
2238
|
-
return {
|
|
2239
|
-
reservation,
|
|
2240
|
-
decision:
|
|
2241
|
-
usage === undefined
|
|
2242
|
-
? { kind: "unresolved", knownUsd: undefined }
|
|
2243
|
-
: settlementFromUsage(usage),
|
|
2244
|
-
};
|
|
2245
|
-
}
|
|
2237
|
+
// but kept total: no CAS transition is applied and no fence is claimed.
|
|
2238
|
+
if (ledger === undefined) return reservation;
|
|
2246
2239
|
const idempotencyKey = `${reservation.reservationId}:finalize`;
|
|
2247
2240
|
const usage =
|
|
2248
2241
|
result.kind === "cancelled" ? undefined : result.outcome.usage;
|
|
2249
2242
|
if (usage === undefined) {
|
|
2250
2243
|
await ledger.markUnresolvedRemote(
|
|
2251
2244
|
reservation.reservationId,
|
|
2252
|
-
undefined,
|
|
2253
2245
|
idempotencyKey,
|
|
2254
2246
|
);
|
|
2255
|
-
return {
|
|
2256
|
-
reservation: {
|
|
2257
|
-
...reservation,
|
|
2258
|
-
state: "unresolved_remote",
|
|
2259
|
-
knownUsd: undefined,
|
|
2260
|
-
},
|
|
2261
|
-
decision: { kind: "unresolved", knownUsd: undefined },
|
|
2262
|
-
};
|
|
2247
|
+
return { ...reservation, state: "unresolved_remote" };
|
|
2263
2248
|
}
|
|
2264
2249
|
const decision = settlementFromUsage(usage);
|
|
2265
2250
|
if (decision.kind === "settle") {
|
|
2266
2251
|
await ledger.settle(reservation.reservationId, decision, idempotencyKey);
|
|
2267
|
-
return {
|
|
2268
|
-
reservation: {
|
|
2269
|
-
...reservation,
|
|
2270
|
-
state: "settled",
|
|
2271
|
-
settledUsd: decision.actualUsd,
|
|
2272
|
-
},
|
|
2273
|
-
decision,
|
|
2274
|
-
};
|
|
2252
|
+
return { ...reservation, state: "settled" };
|
|
2275
2253
|
}
|
|
2276
|
-
// #182 follow-up, free-route fence scope: a free binding reserves
|
|
2277
|
-
//
|
|
2278
|
-
//
|
|
2279
|
-
//
|
|
2280
|
-
//
|
|
2281
|
-
//
|
|
2282
|
-
// that bills nothing there are no dollars to stop spending over.
|
|
2254
|
+
// #182 follow-up, free-route fence scope: a free binding reserves, but
|
|
2255
|
+
// fencing the shared credential bucket on EVERY terminal free outcome
|
|
2256
|
+
// meant one transient failure (parse failure, watchdog timeout shape)
|
|
2257
|
+
// killed the whole all-free run. That contradicts the neighboring
|
|
2258
|
+
// subscription reasoning in production-runtime.ts MultiProviderRunner.run:
|
|
2259
|
+
// a route that bills nothing has no credential worth protecting.
|
|
2283
2260
|
//
|
|
2284
|
-
// So an unresolved free outcome
|
|
2285
|
-
//
|
|
2286
|
-
//
|
|
2287
|
-
//
|
|
2288
|
-
//
|
|
2289
|
-
//
|
|
2290
|
-
//
|
|
2291
|
-
//
|
|
2292
|
-
//
|
|
2293
|
-
// collectUnresolvedSpend (pipeline.ts) and renderResult (ui/result.ts)
|
|
2294
|
-
// only read unresolved_remote reservations, so a released_unstarted free
|
|
2295
|
-
// attempt is invisible to both — no floor marker, no unresolved row.
|
|
2261
|
+
// So an unresolved free outcome releases instead of fencing: a terminal
|
|
2262
|
+
// released_unstarted fences nothing — the same-step retry reserves fine
|
|
2263
|
+
// and sibling steps are unaffected. Coupling-1 is preserved: a non-complete
|
|
2264
|
+
// usage still never settles, it just releases instead of fencing.
|
|
2265
|
+
// Cancelled/undefined-usage keeps the fence (no usage means no evidence
|
|
2266
|
+
// either way — fail closed), and the metered path below is untouched.
|
|
2267
|
+
// Downstream-safe: collectUnresolvedSpend (pipeline.ts) and renderResult
|
|
2268
|
+
// (ui/result.ts) only read unresolved_remote reservations, so a
|
|
2269
|
+
// released_unstarted free attempt is invisible to both.
|
|
2296
2270
|
//
|
|
2297
2271
|
// #187: watchdog timeout on a free route synthesizes an unavailable usage;
|
|
2298
2272
|
// preserving the transport's free billing mode ensures it releases
|
|
2299
2273
|
// unstarted instead of fencing the bucket, allowing attempt 2 to retry.
|
|
2300
2274
|
const isFreeRoute =
|
|
2301
2275
|
usage.billingMode === "free" || transport?.billingMode === "free";
|
|
2302
|
-
if (isFreeRoute
|
|
2276
|
+
if (isFreeRoute) {
|
|
2303
2277
|
await ledger.releaseUnstarted(reservation.reservationId, idempotencyKey);
|
|
2304
|
-
return {
|
|
2305
|
-
reservation: { ...reservation, state: "released_unstarted" },
|
|
2306
|
-
decision,
|
|
2307
|
-
};
|
|
2278
|
+
return { ...reservation, state: "released_unstarted" };
|
|
2308
2279
|
}
|
|
2309
2280
|
await ledger.markUnresolvedRemote(
|
|
2310
2281
|
reservation.reservationId,
|
|
2311
|
-
decision.knownUsd,
|
|
2312
2282
|
idempotencyKey,
|
|
2313
2283
|
);
|
|
2314
|
-
return {
|
|
2315
|
-
reservation: {
|
|
2316
|
-
...reservation,
|
|
2317
|
-
state: "unresolved_remote",
|
|
2318
|
-
knownUsd: decision.knownUsd,
|
|
2319
|
-
},
|
|
2320
|
-
decision,
|
|
2321
|
-
};
|
|
2284
|
+
return { ...reservation, state: "unresolved_remote" };
|
|
2322
2285
|
}
|
|
2323
2286
|
|
|
2324
2287
|
// The body of a single attempt, unchanged from PR0 except for the §9.2
|
|
@@ -2541,7 +2504,16 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
2541
2504
|
if (execution.cancelled) {
|
|
2542
2505
|
// A cancelled attempt has no outcome to carry a warning, but what it did
|
|
2543
2506
|
// before the stop is still worth keeping.
|
|
2544
|
-
await copyClaudeSessionTrace(step, request, attempt);
|
|
2507
|
+
const miss = await copyClaudeSessionTrace(step, request, attempt);
|
|
2508
|
+
await writeAttemptUsageFacts({
|
|
2509
|
+
step,
|
|
2510
|
+
route: request.route,
|
|
2511
|
+
attempt,
|
|
2512
|
+
transport,
|
|
2513
|
+
miss,
|
|
2514
|
+
outcome: undefined,
|
|
2515
|
+
billingMode: this.attemptBillingMode(transport),
|
|
2516
|
+
});
|
|
2545
2517
|
return { kind: "cancelled" };
|
|
2546
2518
|
}
|
|
2547
2519
|
|
|
@@ -2549,12 +2521,18 @@ export class StepExecutionHarness implements StepRunner {
|
|
|
2549
2521
|
// `execution.cancelled` is the only path executeSession takes without
|
|
2550
2522
|
// ever invoking `onData`, and `onData` is what both delivery variants
|
|
2551
2523
|
// above come from.
|
|
2552
|
-
const
|
|
2524
|
+
const settled = execution.outcome as TransportOutcome;
|
|
2525
|
+
const miss = await copyClaudeSessionTrace(step, request, attempt);
|
|
2526
|
+
await writeAttemptUsageFacts({
|
|
2553
2527
|
step,
|
|
2554
|
-
request,
|
|
2528
|
+
route: request.route,
|
|
2555
2529
|
attempt,
|
|
2556
|
-
|
|
2557
|
-
|
|
2530
|
+
transport,
|
|
2531
|
+
miss,
|
|
2532
|
+
outcome: settled,
|
|
2533
|
+
billingMode: settled.usage.billingMode,
|
|
2534
|
+
});
|
|
2535
|
+
const outcome = withTraceMissAnnotation(settled, miss);
|
|
2558
2536
|
|
|
2559
2537
|
if (execution.delivery?.delivered) {
|
|
2560
2538
|
return { kind: "delivered", outcome, parsed: execution.delivery.parsed };
|