crysk 0.7.0 → 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 +27 -23
- package/dist/SKILL-1q3658n0.md +74 -0
- package/dist/SKILL-25j4q9b8.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/cli.js +24658 -24654
- package/dist/index.js +492 -463
- 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/workflow-andv2dvm.yml +158 -0
- package/package.json +1 -1
- package/prompts/default/PROVENANCE.md +1 -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-pr/SKILL.md +1 -1
- package/src/asset-manifest.ts +0 -6
- package/src/assets.ts +1 -4
- package/src/brand.ts +20 -30
- package/src/ci/gates.ts +155 -201
- 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/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 +14 -4
- package/src/model/profiles/claude.ts +1 -2
- package/src/model/profiles/openai.ts +1 -2
- package/src/model/profiles.ts +1 -2
- 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/preflight.ts +15 -12
- package/src/pr/publish-outcome.ts +1 -19
- package/src/pr/review-pr.ts +10 -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/review/boundary.ts +4 -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 +57 -395
- package/src/review/preflight.ts +162 -201
- package/src/review/prompt-set.ts +0 -2
- package/src/review/report.ts +2 -162
- package/src/review/review.ts +16 -65
- package/src/review/route-preflight.ts +3 -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/default/deep-review-parity.md +0 -152
- package/prompts/scout.md +0 -140
- package/src/ci/platform-estimate.ts +0 -111
- package/src/review/scout.ts +0 -323
package/src/review/pipeline.ts
CHANGED
|
@@ -4,7 +4,7 @@
|
|
|
4
4
|
// assembly) deterministic and testable here, and spends model tokens only on
|
|
5
5
|
// the hunts and the refutation themselves.
|
|
6
6
|
|
|
7
|
-
import { existsSync } from "node:fs";
|
|
7
|
+
import { existsSync, readFileSync } from "node:fs";
|
|
8
8
|
import { chmod, mkdir, realpath } from "node:fs/promises";
|
|
9
9
|
import path from "node:path";
|
|
10
10
|
import {
|
|
@@ -136,7 +136,6 @@ import type { InternalCapabilityReport } from "../diversity/admission";
|
|
|
136
136
|
import type { BenchmarkTarget } from "../diversity/identity";
|
|
137
137
|
import {
|
|
138
138
|
assertDiversityLegRoutes,
|
|
139
|
-
assertDiversitySpendUnderCap,
|
|
140
139
|
buildDiversityPipelineRecord,
|
|
141
140
|
type DiversityExecutionContext,
|
|
142
141
|
diversityDebugFromLedger,
|
|
@@ -172,6 +171,10 @@ import {
|
|
|
172
171
|
attemptLogPath,
|
|
173
172
|
settlementReceiptPath,
|
|
174
173
|
} from "../execution/step-artifacts";
|
|
174
|
+
import {
|
|
175
|
+
type AttemptUsageFacts,
|
|
176
|
+
usageFactsPath,
|
|
177
|
+
} from "../execution/usage-facts";
|
|
175
178
|
import type { NormalizedUsage } from "../execution/usage-normalized";
|
|
176
179
|
import { sumNormalizedUsage } from "../execution/usage-normalized";
|
|
177
180
|
import { redactEvidenceText } from "../security/evidence-redaction";
|
|
@@ -195,13 +198,6 @@ import {
|
|
|
195
198
|
NO_REVIEW_INSTRUCTIONS,
|
|
196
199
|
ownerPolicyText,
|
|
197
200
|
} from "./preflight";
|
|
198
|
-
import {
|
|
199
|
-
capScoutLeads,
|
|
200
|
-
renderLeadsBlock,
|
|
201
|
-
type ScoutLead,
|
|
202
|
-
scoutPrompt,
|
|
203
|
-
validateScoutLeads,
|
|
204
|
-
} from "./scout";
|
|
205
201
|
import type { ExcludedPath } from "./size-gate";
|
|
206
202
|
import {
|
|
207
203
|
DEFAULT_STEP_MAX_ATTEMPTS,
|
|
@@ -289,7 +285,7 @@ export interface PipelineInput {
|
|
|
289
285
|
// beside `pipelineTimeoutMs` and is injectable for the same reason: without
|
|
290
286
|
// it every ceiling test would sleep 8.5 real seconds.
|
|
291
287
|
ceilingGraceMs?: number;
|
|
292
|
-
// `bundled` on the summarizer
|
|
288
|
+
// `bundled` on the summarizer and consolidate inputs (set by the
|
|
293
289
|
// product paths' pipeline*Input, #review/route-preflight) marks the engine's
|
|
294
290
|
// own prompt, resolved through the active profile's prompt set: its model
|
|
295
291
|
// and effort then come from the profile, never from its frontmatter
|
|
@@ -298,22 +294,14 @@ export interface PipelineInput {
|
|
|
298
294
|
// Optional engine-owned summary step. It is deliberately outside ReviewSpec:
|
|
299
295
|
// this prompt is bundled with pr-hero, not part of the benchmarked agent set.
|
|
300
296
|
summarizer?: { promptPath: string; model?: string; bundled?: boolean };
|
|
301
|
-
// Optional engine-owned scout step (ROADMAP-DOORDASH M5,
|
|
302
|
-
// `docs/scout-design.md` §3.3). Outside ReviewSpec for the summarizer's
|
|
303
|
-
// reason and one more of its own: `ReviewSpec.role` is
|
|
304
|
-
// `"hunter" | "refuter"` with a runtime guard, so a scout cannot be a spec
|
|
305
|
-
// entry without widening that union — and widening it would re-fingerprint
|
|
306
|
-
// the prompt set, which is exactly what M6's one-variable property forbids.
|
|
307
|
-
// ABSENT means the control pipeline, byte for byte.
|
|
308
|
-
scout?: { promptPath: string; model?: string; bundled?: boolean };
|
|
309
297
|
// Optional engine-owned consolidation pass over the POSTED findings, after
|
|
310
298
|
// the refuter (prompts/consolidate.md, measured on the Martian 50: F1
|
|
311
299
|
// 39.7 -> 42.3, lab bench/METRICS.md 2026-10-27). Opt-in on the engine input
|
|
312
300
|
// so the lab's direct `runPipeline` callers are unchanged; the CLI and PR
|
|
313
301
|
// paths turn it on. ABSENT means no consolidation step and findings exactly
|
|
314
302
|
// as before; pipeline.json still carries a `consolidation` record
|
|
315
|
-
// (`enabled: false`, `status: "skipped"`), the
|
|
316
|
-
//
|
|
303
|
+
// (`enabled: false`, `status: "skipped"`), so the two arms are told apart
|
|
304
|
+
// from the artifact alone.
|
|
317
305
|
consolidate?: { promptPath: string; model?: string; bundled?: boolean };
|
|
318
306
|
// The re-review verifier's own body (prompts/verifier.md), engine-owned.
|
|
319
307
|
// The product's PR path passes it. ABSENT means the verify steps render the
|
|
@@ -324,7 +312,7 @@ export interface PipelineInput {
|
|
|
324
312
|
// Optional engine-owned materiality gate (odd/tasks/materiality-gate.md):
|
|
325
313
|
// an isolated judge reads every finding that would post as blocking and
|
|
326
314
|
// demotes the immaterial ones to advisory. ABSENT means no judge step and
|
|
327
|
-
// findings exactly as before
|
|
315
|
+
// findings exactly as before; pipeline.json still
|
|
328
316
|
// carries a `materiality` record (`enabled: false`, `status: "skipped"`).
|
|
329
317
|
// No `model` seat on purpose: the judge runs only on a model it was
|
|
330
318
|
// measured on (sonnet by default, opus through `models.judge`), and neither
|
|
@@ -397,7 +385,7 @@ export interface PipelineInput {
|
|
|
397
385
|
// config's, as the routing config used to.
|
|
398
386
|
routePlan?: ResolvedRoutePlan;
|
|
399
387
|
models?: ModelsConfig;
|
|
400
|
-
// Item 7: skip hunter
|
|
388
|
+
// Item 7: skip hunter fan-out. A re-review whose restricted
|
|
401
389
|
// delta is empty still classifies and verifies (C6); spawning hunters on
|
|
402
390
|
// an empty patch would bill a first-review for nothing. Absent = run
|
|
403
391
|
// hunters, so every existing caller stays byte-identical.
|
|
@@ -515,18 +503,6 @@ export type PipelineProgressEvent =
|
|
|
515
503
|
// Same contract as hunter-finished's reason.
|
|
516
504
|
reason?: string;
|
|
517
505
|
}
|
|
518
|
-
// The scout is the one AWAITED stage between "the run started" and the
|
|
519
|
-
// first hunter spawn, and M4 measured it at 86-600s. Without a started
|
|
520
|
-
// event that is up to ten minutes of a paid run looking hung — the exact
|
|
521
|
-
// incident the progress tap was built for (see PipelineDeps.onProgress).
|
|
522
|
-
| { kind: "scout-started"; model: string }
|
|
523
|
-
| {
|
|
524
|
-
kind: "scout-finished";
|
|
525
|
-
ok: boolean;
|
|
526
|
-
durationMs: number;
|
|
527
|
-
// Leads DELIVERED to the hunters, post-cap. Absent for a failed step.
|
|
528
|
-
leads?: number;
|
|
529
|
-
}
|
|
530
506
|
// A step about to be retried. Until this existed a retrying hunter looked
|
|
531
507
|
// merely slow: `attempts` was counted in PerAgentUsage after the fact, which
|
|
532
508
|
// is a post-mortem, never a live signal. OBSERVATION ONLY — the runner's
|
|
@@ -616,7 +592,6 @@ export interface PerAgentUsage {
|
|
|
616
592
|
duration_ms: number;
|
|
617
593
|
tokens_in: number;
|
|
618
594
|
tokens_out: number;
|
|
619
|
-
cost_usd_est: number;
|
|
620
595
|
attempts: number;
|
|
621
596
|
status: string;
|
|
622
597
|
// Why a failed step failed, in the provider's own words: the transport's one-line
|
|
@@ -633,17 +608,14 @@ export interface PipelineResult {
|
|
|
633
608
|
usage: SessionUsage;
|
|
634
609
|
// #173: the run-level normalized rollup, the SAME object `pipeline.json`
|
|
635
610
|
// publishes as its `usage_v2` key — handed to the caller so the terminal and
|
|
636
|
-
// report.md can show
|
|
637
|
-
// show cash and notional totals separately"). The legacy `usage` above
|
|
638
|
-
// projects `cashCostUsd` only, and on a Claude subscription route that is a
|
|
639
|
-
// truthful $0.00 with the list-basis figure living here instead.
|
|
611
|
+
// report.md can show the run's tokens and credential class.
|
|
640
612
|
//
|
|
641
613
|
// Taken off `state.usageTotalV2` rather than re-accumulated for the display
|
|
642
614
|
// path: a second tally is the "parallel mechanism able to disagree with
|
|
643
615
|
// per_agent" the per-step usage join site already refuses. Absent — never
|
|
644
616
|
// zeroed — when no step reported normalized usage, because the display
|
|
645
|
-
// surfaces key on `undefined` to omit their
|
|
646
|
-
// zero would print a
|
|
617
|
+
// surfaces key on `undefined` to omit their usage row, and a fabricated
|
|
618
|
+
// zero would print a claim for a run that produced none.
|
|
647
619
|
usageV2?: NormalizedUsage;
|
|
648
620
|
// True only when ALL hunter steps failed — nothing was hunted. A single
|
|
649
621
|
// hunter failure is a partial run (prose Step 4: "proceed with whatever
|
|
@@ -948,17 +920,10 @@ const SUMMARY_OUTPUT_CONTRACT = [
|
|
|
948
920
|
"The score_reason is 1-2 sentences explaining that advisory score.",
|
|
949
921
|
].join("\n");
|
|
950
922
|
|
|
951
|
-
//
|
|
952
|
-
// nothing — not a separator, not a blank line (§3.8). That is what makes M6's
|
|
953
|
-
// control arm a control arm: scout off, and scout on returning zero leads,
|
|
954
|
-
// must both produce the byte-identical prompt an unled run produces. C4's
|
|
955
|
-
// wrapBlock preserves that property by returning "" for empty content, so an
|
|
956
|
-
// empty leads block still leaves no trace — not even an empty tag pair.
|
|
957
|
-
//
|
|
958
|
-
// Both non-engine blocks here are wrapped in the run's nonced boundary tags
|
|
923
|
+
// The non-engine blocks here are wrapped in the run's nonced boundary tags
|
|
959
924
|
// (`docs/c4-preamble-design.md` §3.3-§3.4): the patch is the PR author's, and
|
|
960
|
-
//
|
|
961
|
-
//
|
|
925
|
+
// so is the intent. `nonce` sits ahead of the defaulted parameter so a caller
|
|
926
|
+
// cannot reach this composer without one.
|
|
962
927
|
// The codegraph mandate, appended only when the run's MCP registry grants
|
|
963
928
|
// codegraph (`codegraphGranted`).
|
|
964
929
|
//
|
|
@@ -1012,7 +977,7 @@ async function codegraphGranted(mcpConfigPath: string): Promise<boolean> {
|
|
|
1012
977
|
//
|
|
1013
978
|
// WHY the hunters and nobody else: the refuter judges a claim against the code
|
|
1014
979
|
// alone, so that an author cannot talk it out of a finding by saying the
|
|
1015
|
-
// behaviour is intentional. The summarizer,
|
|
980
|
+
// behaviour is intentional. The summarizer, consolidation and
|
|
1016
981
|
// adjudicator work from the diff or the findings and never see it either.
|
|
1017
982
|
//
|
|
1018
983
|
// WHY it renders to NOTHING when both are absent: a hunter's request without
|
|
@@ -1086,10 +1051,8 @@ function hunterPrompt(
|
|
|
1086
1051
|
nonce: string,
|
|
1087
1052
|
toolsEnabled: boolean,
|
|
1088
1053
|
codegraph: boolean,
|
|
1089
|
-
leadsBlock = "",
|
|
1090
1054
|
intent: PrIntent = { title: "", description: "" },
|
|
1091
1055
|
): string {
|
|
1092
|
-
const wrappedLeads = wrapBlock("scout_leads", nonce, leadsBlock);
|
|
1093
1056
|
const wrappedTitle = wrapBlock("pr_title", nonce, intent.title);
|
|
1094
1057
|
const wrappedDescription = wrapBlock(
|
|
1095
1058
|
"pr_description",
|
|
@@ -1118,9 +1081,6 @@ function hunterPrompt(
|
|
|
1118
1081
|
"`hops_used` and `hop_trail` are self-reported and may be cross-checked",
|
|
1119
1082
|
"against this run's telemetry MCP-call counts.",
|
|
1120
1083
|
"",
|
|
1121
|
-
// Leads sit LAST before the contract so the diff is still what the hunter
|
|
1122
|
-
// reads first (§3.8's block order).
|
|
1123
|
-
...(wrappedLeads.length === 0 ? [] : [wrappedLeads, ""]),
|
|
1124
1084
|
hunterOutputContract(toolsEnabled),
|
|
1125
1085
|
].join("\n");
|
|
1126
1086
|
}
|
|
@@ -1234,12 +1194,12 @@ function agentRoleEffort(input: PipelineInput, agent: AgentSpec): Effort {
|
|
|
1234
1194
|
// profile's route for its role, never its frontmatter (S3), so the step
|
|
1235
1195
|
// agrees with the plan (buildResolvedRoutePlan's `bundled`). Unmarked, its
|
|
1236
1196
|
// frontmatter holds both seats; `profileFallback` says whether the profile
|
|
1237
|
-
// still backs an absent `model:` there (
|
|
1197
|
+
// still backs an absent `model:` there (consolidate) or not
|
|
1238
1198
|
// (summarizer, which has always failed without one).
|
|
1239
1199
|
function enginePromptChoice(
|
|
1240
1200
|
input: PipelineInput,
|
|
1241
1201
|
prompt: { bundled?: boolean },
|
|
1242
|
-
role: "summary" | "
|
|
1202
|
+
role: "summary" | "consolidate",
|
|
1243
1203
|
parsed: ParsedAgent,
|
|
1244
1204
|
profileFallback: boolean,
|
|
1245
1205
|
): { model?: string; effort?: Effort } {
|
|
@@ -1308,18 +1268,17 @@ interface StepMeta {
|
|
|
1308
1268
|
outPath: string;
|
|
1309
1269
|
// D1-10c. Stamped at CONSTRUCTION as "unsettled" and overwritten when the
|
|
1310
1270
|
// step settles, so every row in a v1.0.0 plan answers "did this run" without
|
|
1311
|
-
// the reader inferring anything from an absent key —
|
|
1312
|
-
// ("the key is missing" is indistinguishable from "this run predates
|
|
1313
|
-
// key"), applied to the eight steps that never had it.
|
|
1271
|
+
// the reader inferring anything from an absent key — ConsolidationRecord's rule
|
|
1272
|
+
// below ("the key is missing" is indistinguishable from "this run predates
|
|
1273
|
+
// the key"), applied to the eight steps that never had it.
|
|
1314
1274
|
//
|
|
1315
1275
|
// WHY "unsettled" and not "not-run" or "skipped":
|
|
1316
1276
|
// - "not-run" would LIE about the commonest case. When the pipeline ceiling
|
|
1317
1277
|
// fires, the steps in flight had already been spawned and had already
|
|
1318
1278
|
// burned tokens; they simply never came back before the snapshot. A row
|
|
1319
1279
|
// claiming they never ran contradicts their own attempt logs on disk.
|
|
1320
|
-
// - "skipped" is taken,
|
|
1321
|
-
//
|
|
1322
|
-
// for". Two neighbouring records spelling one word two ways is how a
|
|
1280
|
+
// - "skipped" is taken, below, and means something else:
|
|
1281
|
+
// ConsolidationRecord uses it for "the stage was never asked for". Two neighbouring records spelling one word two ways is how a
|
|
1323
1282
|
// reader learns the wrong meaning.
|
|
1324
1283
|
// "unsettled" is true of both reachable cases — never spawned, and spawned
|
|
1325
1284
|
// but abandoned — and claims nothing beyond "no terminal verdict reached
|
|
@@ -1348,6 +1307,10 @@ interface StepMeta {
|
|
|
1348
1307
|
// relative to the run dir. Absent when none were copied, so a failed
|
|
1349
1308
|
// capture names no file and a pipeline.json from before traces still parses.
|
|
1350
1309
|
tracePaths?: string[];
|
|
1310
|
+
// platform-owned-cost T6: what each attempt measured per model request, as the
|
|
1311
|
+
// harness wrote it (execution/usage-facts.ts), in attempt order. The platform
|
|
1312
|
+
// report copies it into its `usage_facts`. Absent when no attempt wrote one.
|
|
1313
|
+
usage_facts?: AttemptUsageFacts[];
|
|
1351
1314
|
// D1-08 PR2 (§8): the step's normalized usage, attached by recordSettlement
|
|
1352
1315
|
// — the single existing usage join site (D1-10c's own comment: a second
|
|
1353
1316
|
// pass over `state.steps` "would be a parallel mechanism able to disagree
|
|
@@ -1387,34 +1350,8 @@ interface StepMeta {
|
|
|
1387
1350
|
}>;
|
|
1388
1351
|
}
|
|
1389
1352
|
|
|
1390
|
-
// pipeline.json's `
|
|
1391
|
-
//
|
|
1392
|
-
// artifact alone, and "the key is missing" is indistinguishable from "this run
|
|
1393
|
-
// predates the key".
|
|
1394
|
-
interface ScoutRecord {
|
|
1395
|
-
enabled: boolean;
|
|
1396
|
-
model?: string;
|
|
1397
|
-
// "skipped" is the flag being off. "failed" and "ok" both mean the stage
|
|
1398
|
-
// was asked for — a failed scout is a CONTROL-arm run wearing a scout-arm
|
|
1399
|
-
// flag, and §3.6's M6 rule excludes and re-runs it on exactly this field.
|
|
1400
|
-
status: "ok" | "failed" | "skipped";
|
|
1401
|
-
// DELIVERED leads, post-cap — the number the hunters actually saw.
|
|
1402
|
-
leads_count: number;
|
|
1403
|
-
// Leads the caps dropped. A truncation that fires routinely is a PROMPT
|
|
1404
|
-
// defect to fix, never a cap to raise (§3.8), so it is recorded per run
|
|
1405
|
-
// rather than left to a probe nobody re-runs.
|
|
1406
|
-
leads_truncated: number;
|
|
1407
|
-
// `why` sentences the 240-char cap cut. Additive beyond §3.9's list, and
|
|
1408
|
-
// deliberately: M5 INHERITS this defect from M4 (it fired in most runs), and
|
|
1409
|
-
// a defect nothing counts in production is a defect nobody notices.
|
|
1410
|
-
why_truncated: number;
|
|
1411
|
-
prompt_sha256?: string;
|
|
1412
|
-
duration_ms: number;
|
|
1413
|
-
}
|
|
1414
|
-
|
|
1415
|
-
// pipeline.json's `consolidation` key, written on every run for ScoutRecord's
|
|
1416
|
-
// reason: "the key is missing" is indistinguishable from "this run predates
|
|
1417
|
-
// the key". "skipped" means the stage was not asked for (or the ceiling cut
|
|
1353
|
+
// pipeline.json's `consolidation` key, written on every run: "the key is
|
|
1354
|
+
// missing" is indistinguishable from "this run predates the key". "skipped" means the stage was not asked for (or the ceiling cut
|
|
1418
1355
|
// the run first); "passthrough" means it was asked for and had nothing it
|
|
1419
1356
|
// could safely do (no posted findings, or a claim carrying the nonce), so it
|
|
1420
1357
|
// never spawned; "failed" means it was asked for and its plan was not applied.
|
|
@@ -1433,8 +1370,8 @@ interface ConsolidationRecord {
|
|
|
1433
1370
|
duration_ms: number;
|
|
1434
1371
|
}
|
|
1435
1372
|
|
|
1436
|
-
// pipeline.json's `materiality` key, written on every run for
|
|
1437
|
-
// reason: "the key is missing" is indistinguishable from "this run predates
|
|
1373
|
+
// pipeline.json's `materiality` key, written on every run for
|
|
1374
|
+
// ConsolidationRecord's reason: "the key is missing" is indistinguishable from "this run predates
|
|
1438
1375
|
// the key". "skipped" means the gate was not asked for; "ceiling" means it was
|
|
1439
1376
|
// asked for and the pipeline ceiling (or its budget) stood it down before it
|
|
1440
1377
|
// spawned; "unsupported_route" means the judge would not have run the
|
|
@@ -1473,7 +1410,6 @@ interface RunState {
|
|
|
1473
1410
|
// after the hunters', and a captured undefined would silently disable the
|
|
1474
1411
|
// check on exactly the leg that motivated the issue.
|
|
1475
1412
|
resolveProofRef?: (candidate: string) => boolean;
|
|
1476
|
-
scout?: ScoutRecord;
|
|
1477
1413
|
consolidation?: ConsolidationRecord;
|
|
1478
1414
|
// Set only by a consolidation step that returned a VALID plan. finish()
|
|
1479
1415
|
// applies it; the ceiling path reaches finish() without one and so never
|
|
@@ -1568,7 +1504,7 @@ interface RunState {
|
|
|
1568
1504
|
// D1-08 PR2: the run-level rollup of every step's normalized usage, summed
|
|
1569
1505
|
// via `accumulateUsageV2` alongside each existing `usageTotal` accumulation
|
|
1570
1506
|
// site. `undefined` until the first step reports `usageV2` — seeding it
|
|
1571
|
-
// with a billingMode
|
|
1507
|
+
// with a billingMode-less zero would force `sumNormalizedUsage`
|
|
1572
1508
|
// to collapse the run's real billing mode to "unknown" on the first sum.
|
|
1573
1509
|
usageTotalV2?: NormalizedUsage;
|
|
1574
1510
|
steps: StepMeta[];
|
|
@@ -1828,15 +1764,6 @@ async function execute(
|
|
|
1828
1764
|
// ignore if not readable
|
|
1829
1765
|
}
|
|
1830
1766
|
}
|
|
1831
|
-
let scoutFrontmatter: string | undefined;
|
|
1832
|
-
if (input.scout && input.scout.bundled !== true) {
|
|
1833
|
-
try {
|
|
1834
|
-
const parsed = await parseAgentFile(input.scout.promptPath);
|
|
1835
|
-
scoutFrontmatter = parsed.model;
|
|
1836
|
-
} catch {
|
|
1837
|
-
// ignore if not readable
|
|
1838
|
-
}
|
|
1839
|
-
}
|
|
1840
1767
|
let consolidateFrontmatter: string | undefined;
|
|
1841
1768
|
if (input.consolidate && input.consolidate.bundled !== true) {
|
|
1842
1769
|
try {
|
|
@@ -1863,15 +1790,6 @@ async function execute(
|
|
|
1863
1790
|
bundled: input.summarizer.bundled === true,
|
|
1864
1791
|
},
|
|
1865
1792
|
}),
|
|
1866
|
-
...(input.scout === undefined
|
|
1867
|
-
? {}
|
|
1868
|
-
: {
|
|
1869
|
-
scout: {
|
|
1870
|
-
model: input.scout.model,
|
|
1871
|
-
frontmatterModel: scoutFrontmatter,
|
|
1872
|
-
bundled: input.scout.bundled === true,
|
|
1873
|
-
},
|
|
1874
|
-
}),
|
|
1875
1793
|
...(input.consolidate === undefined
|
|
1876
1794
|
? {}
|
|
1877
1795
|
: {
|
|
@@ -1977,8 +1895,7 @@ async function execute(
|
|
|
1977
1895
|
|
|
1978
1896
|
// Step 3a — the run's boundary nonce (C4 O-3.3). HERE, once, and identical
|
|
1979
1897
|
// for every step of this run: a per-step nonce would give each step a
|
|
1980
|
-
// different prompt for the same bytes, and
|
|
1981
|
-
// comparison. Drawn against exactly the blocks that already exist — the
|
|
1898
|
+
// different prompt for the same bytes, and runs are compared byte for byte. Drawn against exactly the blocks that already exist — the
|
|
1982
1899
|
// patch, the operator gotchas, the operator priors, the author's intent —
|
|
1983
1900
|
// because those are the only ones whose content is fixed before the draw.
|
|
1984
1901
|
// `renderPriorsBlock` is called here and again inside `renderAgentBody` on
|
|
@@ -2004,19 +1921,6 @@ async function execute(
|
|
|
2004
1921
|
|
|
2005
1922
|
const stepsDir = path.join(input.runDir, "steps");
|
|
2006
1923
|
|
|
2007
|
-
// Step 3b — the scout (ROADMAP-DOORDASH M5). Here and not elsewhere: after
|
|
2008
|
-
// `patch` and the trigger decision exist, strictly before the hunter
|
|
2009
|
-
// composition loop that consumes its leads. It is AWAITED, unlike the
|
|
2010
|
-
// summarizer, which puts it on the critical path — the cost this design
|
|
2011
|
-
// states out loud (§3.9) rather than hiding.
|
|
2012
|
-
// A ceiling that fired before the scout was even spawned skips it and hunts
|
|
2013
|
-
// without leads — the control pipeline's shape, which is a degraded run, not
|
|
2014
|
-
// a broken one.
|
|
2015
|
-
const leadsBlock =
|
|
2016
|
-
skipDiscovery || ceilingAborted(deps, state)
|
|
2017
|
-
? ""
|
|
2018
|
-
: await runScout(input, deps, state, patch, stepsDir, boundaryNonce);
|
|
2019
|
-
|
|
2020
1924
|
// Step 4 — hunter fan-out.
|
|
2021
1925
|
const stepTimeoutMs = input.stepTimeoutMs ?? DEFAULT_STEP_TIMEOUT_MS;
|
|
2022
1926
|
// §5.3 admission, checked BEFORE the composition loop and not at the join:
|
|
@@ -2053,7 +1957,6 @@ async function execute(
|
|
|
2053
1957
|
boundaryNonce,
|
|
2054
1958
|
agent.tools.length > 0,
|
|
2055
1959
|
codegraph && agent.tools.includes(CODEGRAPH_TOOL),
|
|
2056
|
-
leadsBlock,
|
|
2057
1960
|
prIntent,
|
|
2058
1961
|
);
|
|
2059
1962
|
// The rendered body is written run-dir-side as an audit artifact: the
|
|
@@ -2454,7 +2357,6 @@ async function execute(
|
|
|
2454
2357
|
),
|
|
2455
2358
|
};
|
|
2456
2359
|
state.diversity = diversityCtx;
|
|
2457
|
-
assertDiversitySpendUnderCap(plan, diversityCtx.ledger);
|
|
2458
2360
|
}
|
|
2459
2361
|
if (result.status !== "ok") {
|
|
2460
2362
|
state.hunterFailures++;
|
|
@@ -2897,7 +2799,7 @@ async function runRefuter(
|
|
|
2897
2799
|
state.verdicts.set(id, "inconclusive");
|
|
2898
2800
|
}
|
|
2899
2801
|
// The spec carries ONE refuter agent, so its telemetry stays one row —
|
|
2900
|
-
// token
|
|
2802
|
+
// token fields summed across the steps it fanned into, which keeps
|
|
2901
2803
|
// `per_agent` totals reconcilable against the run total. `duration_ms` is the
|
|
2902
2804
|
// one field that cannot be summed: it reports the leg's measured elapsed
|
|
2903
2805
|
// time. `attempts` IS a sum on purpose — N steps have no single attempt
|
|
@@ -2909,7 +2811,6 @@ async function runRefuter(
|
|
|
2909
2811
|
duration_ms: legElapsedMs,
|
|
2910
2812
|
tokens_in: usage.tokens_in,
|
|
2911
2813
|
tokens_out: usage.tokens_out,
|
|
2912
|
-
cost_usd_est: usage.cost_usd_est,
|
|
2913
2814
|
attempts,
|
|
2914
2815
|
status: anyFailed ? "failed" : "ok",
|
|
2915
2816
|
...(anyFailed ? firstFailureDetail(settled) : {}),
|
|
@@ -3104,7 +3005,6 @@ async function runVerify(
|
|
|
3104
3005
|
duration_ms: legElapsedMs,
|
|
3105
3006
|
tokens_in: usage.tokens_in,
|
|
3106
3007
|
tokens_out: usage.tokens_out,
|
|
3107
|
-
cost_usd_est: usage.cost_usd_est,
|
|
3108
3008
|
attempts,
|
|
3109
3009
|
status: anyFailed ? "failed" : "ok",
|
|
3110
3010
|
...(anyFailed ? firstFailureDetail(settled) : {}),
|
|
@@ -3258,7 +3158,6 @@ async function runAdjudicate(
|
|
|
3258
3158
|
duration_ms: legElapsedMs,
|
|
3259
3159
|
tokens_in: usage.tokens_in,
|
|
3260
3160
|
tokens_out: usage.tokens_out,
|
|
3261
|
-
cost_usd_est: usage.cost_usd_est,
|
|
3262
3161
|
attempts,
|
|
3263
3162
|
status: anyFailed ? "failed" : "ok",
|
|
3264
3163
|
...(anyFailed ? firstFailureDetail(settled) : {}),
|
|
@@ -3268,251 +3167,9 @@ async function runAdjudicate(
|
|
|
3268
3167
|
return outcomes;
|
|
3269
3168
|
}
|
|
3270
3169
|
|
|
3271
|
-
// The scout's non-hunter budgets (§3.5 mechanism 4), and the one number here
|
|
3272
|
-
// that departs from the ratified design: §3.5 says "a 5-minute watchdog",
|
|
3273
|
-
// copying the summarizer. M4 then MEASURED the stage at 86-600s across 60
|
|
3274
|
-
// spawns and raised its own probe watchdog from 10 to 15 minutes on that
|
|
3275
|
-
// evidence (§3.10bis). A 5-minute ceiling would reap runs the only data we
|
|
3276
|
-
// have calls normal, so the measurement wins over the estimate that preceded
|
|
3277
|
-
// it. One attempt is unchanged: a scout that could not answer once has
|
|
3278
|
-
// nothing the run needs badly enough to pay twice for.
|
|
3279
|
-
const SCOUT_TIMEOUT_MS = 15 * 60 * 1000;
|
|
3280
|
-
|
|
3281
|
-
// Last resort in the model chain, and it exists because `prompts/scout.md`
|
|
3282
|
-
// deliberately carries NO `model:` frontmatter: that file is M4's ratified
|
|
3283
|
-
// artifact (sha256 68a81d26081e, v5), and adding a line to it would move the
|
|
3284
|
-
// sha the milestone recorded for a prompt whose body nothing changed.
|
|
3285
|
-
//
|
|
3286
|
-
// The value is not a taste call either. M4 ran every one of its 60 measured
|
|
3287
|
-
// spawns on sonnet (`scripts/scout-probe.ts` defaults to it and never reads
|
|
3288
|
-
// the prompt's frontmatter), and M6's whole control corpus is sonnet (§1.2) —
|
|
3289
|
-
// so this is the model the scout has actually been measured on, and the model
|
|
3290
|
-
// the A/B will run it on. `--scout-model` is the documented exit; the cheap
|
|
3291
|
-
// tier is its own later experiment (§3.13), not a variable inside M6. Read
|
|
3292
|
-
// from the `claude` profile, which holds the value now (model-profiles S2).
|
|
3293
|
-
export const DEFAULT_SCOUT_MODEL = profileModelAlias(
|
|
3294
|
-
profileRoute(FALLBACK_PROFILE, "scout").model,
|
|
3295
|
-
);
|
|
3296
|
-
|
|
3297
|
-
// Runs the scout and returns the block to append to every hunter prompt —
|
|
3298
|
-
// the EMPTY STRING whenever there is nothing to append, which covers all four
|
|
3299
|
-
// quiet paths: the flag is off, the prompt file would not parse, the step
|
|
3300
|
-
// failed, and the scout honestly found nothing.
|
|
3301
|
-
//
|
|
3302
|
-
// FAIL-OPEN, and it is the load-bearing property (§3.6): nothing in here sets
|
|
3303
|
-
// `state.partial`. A run without a scout is the CONTROL pipeline, which is by
|
|
3304
|
-
// definition complete — it cannot have lost a finding it was never going to
|
|
3305
|
-
// produce. #42's incompleteness notice is about a review that lost a hunter or
|
|
3306
|
-
// the refuter, which is a different thing and stays untouched. What this must
|
|
3307
|
-
// never be is SILENT, so the failure lands in `pipeline.json` (`status`), in
|
|
3308
|
-
// `per_agent.scout`, and on the progress tap.
|
|
3309
|
-
async function runScout(
|
|
3310
|
-
input: PipelineInput,
|
|
3311
|
-
deps: PipelineDeps,
|
|
3312
|
-
state: RunState,
|
|
3313
|
-
patch: string,
|
|
3314
|
-
stepsDir: string,
|
|
3315
|
-
nonce: string,
|
|
3316
|
-
): Promise<string> {
|
|
3317
|
-
if (!input.scout) {
|
|
3318
|
-
state.scout = {
|
|
3319
|
-
enabled: false,
|
|
3320
|
-
status: "skipped",
|
|
3321
|
-
leads_count: 0,
|
|
3322
|
-
leads_truncated: 0,
|
|
3323
|
-
why_truncated: 0,
|
|
3324
|
-
duration_ms: 0,
|
|
3325
|
-
};
|
|
3326
|
-
return "";
|
|
3327
|
-
}
|
|
3328
|
-
|
|
3329
|
-
const name = "scout";
|
|
3330
|
-
const systemPromptPath = path.join(stepsDir, `${name}.system.md`);
|
|
3331
|
-
// The RAW validated leads, pre-cap, beside every other step's output — the
|
|
3332
|
-
// runner writes it. Capping happens in the driver afterwards, so the
|
|
3333
|
-
// artifact records what the scout SAID and pipeline.json records what the
|
|
3334
|
-
// hunters were given; collapsing the two would delete the evidence that a
|
|
3335
|
-
// cap fired at all.
|
|
3336
|
-
const outPath = path.join(stepsDir, `${name}.leads.json`);
|
|
3337
|
-
const meta: StepMeta = {
|
|
3338
|
-
name,
|
|
3339
|
-
model: input.scout.model ?? input.model ?? "unresolved",
|
|
3340
|
-
effort: profileRoleEffort(activeProfile(input), "scout"),
|
|
3341
|
-
// FORCED to empty here, never read from the prompt file's frontmatter.
|
|
3342
|
-
// §3.5 mechanism 1 — "the scout cannot open a file, grep, or walk a call
|
|
3343
|
-
// graph" — is the guarantee this whole design rests on, and a guarantee a
|
|
3344
|
-
// prompt edit can revoke is not a guarantee. `model:` in that frontmatter
|
|
3345
|
-
// is still honoured; `tools:` is not.
|
|
3346
|
-
tools: [],
|
|
3347
|
-
systemPromptPath,
|
|
3348
|
-
outPath,
|
|
3349
|
-
// Same construction-time default `stepMeta()` stamps; this meta is built by
|
|
3350
|
-
// hand because the scout resolves its model and tools as it goes.
|
|
3351
|
-
status: "unsettled",
|
|
3352
|
-
};
|
|
3353
|
-
const record: ScoutRecord = {
|
|
3354
|
-
enabled: true,
|
|
3355
|
-
model: meta.model,
|
|
3356
|
-
// Pessimistic default: every early return below is a failure, and only
|
|
3357
|
-
// the delivered path overwrites it.
|
|
3358
|
-
status: "failed",
|
|
3359
|
-
leads_count: 0,
|
|
3360
|
-
leads_truncated: 0,
|
|
3361
|
-
why_truncated: 0,
|
|
3362
|
-
duration_ms: 0,
|
|
3363
|
-
};
|
|
3364
|
-
state.scout = record;
|
|
3365
|
-
const startedAt = Date.now();
|
|
3366
|
-
|
|
3367
|
-
const abandon = (error?: unknown): string => {
|
|
3368
|
-
// Status only. `attempts` and the two pointers are NOT cleared here on
|
|
3369
|
-
// purpose: `abandon` is reached by three different roads — a prompt file
|
|
3370
|
-
// that never parsed, a runner that threw, and a step that SETTLED and then
|
|
3371
|
-
// failed (or delivered leads that forged the nonce). On that third road a
|
|
3372
|
-
// paid attempt really ran and really wrote its log and receipt, and those
|
|
3373
|
-
// are already recorded by the time this runs. Overwriting them with
|
|
3374
|
-
// "nothing happened" would erase the only evidence separating a scout that
|
|
3375
|
-
// burned money from one that never spawned. A pre-attempt failure has
|
|
3376
|
-
// none of those artifacts, so the reason lands on the step itself (#199)
|
|
3377
|
-
// and is not written over a reason recordSettlement already kept.
|
|
3378
|
-
if (
|
|
3379
|
-
(meta.attempts === undefined || meta.attempts < 1) &&
|
|
3380
|
-
meta.failure === undefined
|
|
3381
|
-
) {
|
|
3382
|
-
recordStepFailure(meta, error ?? "scout failed before its first attempt");
|
|
3383
|
-
} else {
|
|
3384
|
-
meta.status = "failed";
|
|
3385
|
-
}
|
|
3386
|
-
state.perAgent.scout ??= failedAgentEntry();
|
|
3387
|
-
record.duration_ms = Date.now() - startedAt;
|
|
3388
|
-
emit(deps, {
|
|
3389
|
-
kind: "scout-finished",
|
|
3390
|
-
ok: false,
|
|
3391
|
-
durationMs: record.duration_ms,
|
|
3392
|
-
});
|
|
3393
|
-
return "";
|
|
3394
|
-
};
|
|
3395
|
-
|
|
3396
|
-
let spec: StepSpec;
|
|
3397
|
-
try {
|
|
3398
|
-
// Read once, hashed and parsed from the same bytes: a fingerprint over a
|
|
3399
|
-
// second read could disagree with the prompt that actually ran.
|
|
3400
|
-
const raw = await Bun.file(input.scout.promptPath).text();
|
|
3401
|
-
const agent = parseAgentSource(raw);
|
|
3402
|
-
await writeSystemPrompt(systemPromptPath, agent.body);
|
|
3403
|
-
// Precedence, the JD rule extended one seat: --model > --scout-model >
|
|
3404
|
-
// the prompt's frontmatter > the active profile's scout. The profile sits
|
|
3405
|
-
// LAST so a frontmatter model added later still outranks it, unless the
|
|
3406
|
-
// prompt is the engine's own (`bundled`), whose seat is the profile's.
|
|
3407
|
-
const own = enginePromptChoice(input, input.scout, "scout", agent, true);
|
|
3408
|
-
meta.model = resolveModel(
|
|
3409
|
-
input,
|
|
3410
|
-
input.scout.model,
|
|
3411
|
-
own.model,
|
|
3412
|
-
input.scout.promptPath,
|
|
3413
|
-
);
|
|
3414
|
-
record.model = meta.model;
|
|
3415
|
-
const scoutEffort = stepEffortFields(
|
|
3416
|
-
state.routePlan,
|
|
3417
|
-
name,
|
|
3418
|
-
own.effort,
|
|
3419
|
-
profileRoleEffort(activeProfile(input), "scout"),
|
|
3420
|
-
);
|
|
3421
|
-
meta.effort = scoutEffort.effort;
|
|
3422
|
-
// FULL sha256, not the 12 chars the probe prints: the probe's artifact
|
|
3423
|
-
// stores the full digest too, so a pipeline.json and a scout-probe.json
|
|
3424
|
-
// naming the same prompt say the same string.
|
|
3425
|
-
record.prompt_sha256 = new Bun.CryptoHasher("sha256")
|
|
3426
|
-
.update(raw)
|
|
3427
|
-
.digest("hex");
|
|
3428
|
-
spec = {
|
|
3429
|
-
name,
|
|
3430
|
-
systemPromptPath,
|
|
3431
|
-
prompt: scoutPrompt(patch, nonce),
|
|
3432
|
-
tools: meta.tools,
|
|
3433
|
-
mcpConfigPath: input.mcpConfigPath,
|
|
3434
|
-
model: meta.model,
|
|
3435
|
-
...scoutEffort,
|
|
3436
|
-
cwd: input.worktree,
|
|
3437
|
-
outPath,
|
|
3438
|
-
timeoutMs: SCOUT_TIMEOUT_MS,
|
|
3439
|
-
maxAttempts: 1,
|
|
3440
|
-
...stepRouteFields(state.routePlan, name),
|
|
3441
|
-
parse: (finalText) => {
|
|
3442
|
-
const extracted = extractJsonObject(finalText);
|
|
3443
|
-
if (extracted === undefined) {
|
|
3444
|
-
throw new Error("scout final message has no JSON object");
|
|
3445
|
-
}
|
|
3446
|
-
return validateScoutLeads(extracted);
|
|
3447
|
-
},
|
|
3448
|
-
onRetry: (info) => emit(deps, { kind: "step-retry", ...info }),
|
|
3449
|
-
};
|
|
3450
|
-
} catch (error) {
|
|
3451
|
-
// A missing or malformed prompt file. The step still appears in the plan
|
|
3452
|
-
// with status "failed": a stage that was asked for and never spawned must
|
|
3453
|
-
// be visible, or the artifact reads like the flag was off.
|
|
3454
|
-
state.steps.push(meta);
|
|
3455
|
-
return abandon(error);
|
|
3456
|
-
}
|
|
3457
|
-
|
|
3458
|
-
if (spec.route !== undefined) {
|
|
3459
|
-
meta.route = spec.route;
|
|
3460
|
-
Object.assign(meta, routeEffortFields(spec.route));
|
|
3461
|
-
const routedModel = effectiveExecutionModel(spec);
|
|
3462
|
-
meta.model = routedModel;
|
|
3463
|
-
record.model = routedModel;
|
|
3464
|
-
}
|
|
3465
|
-
state.steps.push(meta);
|
|
3466
|
-
emit(deps, { kind: "scout-started", model: meta.model });
|
|
3467
|
-
|
|
3468
|
-
let result: StepResult;
|
|
3469
|
-
try {
|
|
3470
|
-
result = await deps.runner.run(spec);
|
|
3471
|
-
} catch (error) {
|
|
3472
|
-
return abandon(error);
|
|
3473
|
-
}
|
|
3474
|
-
// Usage lands in BOTH seats whatever the verdict — a failed step still
|
|
3475
|
-
// burned tokens, and a run whose bill excludes them under-reports the arm's
|
|
3476
|
-
// cost, which is one of the numbers M6 exists to compare.
|
|
3477
|
-
state.perAgent.scout = perAgentEntry(result);
|
|
3478
|
-
// BEFORE the verdict branches below, so every road that ends in `abandon`
|
|
3479
|
-
// still carries the attempt count and pointers of the session that ran.
|
|
3480
|
-
recordSettlement(meta, result, input.runDir);
|
|
3481
|
-
// The model that ran, when the scout fell back to another (S4).
|
|
3482
|
-
record.model = meta.model;
|
|
3483
|
-
state.usageTotal = sumUsage(state.usageTotal, result.usage);
|
|
3484
|
-
accumulateUsageV2(state, result);
|
|
3485
|
-
if (result.status !== "ok") return abandon();
|
|
3486
|
-
|
|
3487
|
-
const capped = capScoutLeads(result.output as ScoutLead[]);
|
|
3488
|
-
const block = renderLeadsBlock(capped.leads);
|
|
3489
|
-
// The leads are written by a model that READ the nonce in its own prompt, so
|
|
3490
|
-
// a `why` sentence quoting the patch's boundary tag back at us is the real
|
|
3491
|
-
// way this trips — not a random 1-in-2^32 draw. The nonce cannot be
|
|
3492
|
-
// re-drawn here (the scout's paid spawn already carried it, and every step
|
|
3493
|
-
// of a run must share one), so the leads are dropped and the run continues
|
|
3494
|
-
// as the unled CONTROL pipeline. Routed through `abandon` on purpose: the
|
|
3495
|
-
// hunters were not led, which is exactly what a failed scout means, and
|
|
3496
|
-
// `abandon` is what already records that in all three places.
|
|
3497
|
-
if (blockForgesNonce(block, nonce)) return abandon();
|
|
3498
|
-
meta.status = "ok";
|
|
3499
|
-
record.status = "ok";
|
|
3500
|
-
record.leads_count = capped.leads.length;
|
|
3501
|
-
record.leads_truncated = capped.dropped;
|
|
3502
|
-
record.why_truncated = capped.whyTruncated;
|
|
3503
|
-
record.duration_ms = Date.now() - startedAt;
|
|
3504
|
-
emit(deps, {
|
|
3505
|
-
kind: "scout-finished",
|
|
3506
|
-
ok: true,
|
|
3507
|
-
durationMs: record.duration_ms,
|
|
3508
|
-
leads: capped.leads.length,
|
|
3509
|
-
});
|
|
3510
|
-
return renderLeadsBlock(capped.leads);
|
|
3511
|
-
}
|
|
3512
|
-
|
|
3513
3170
|
// The consolidator's budgets, copied from the measurement that justified the
|
|
3514
3171
|
// stage (scripts/consolidate-replay.ts): one claims-only call, so ten minutes
|
|
3515
|
-
// is generous, and two attempts because
|
|
3172
|
+
// is generous, and two attempts because a lost answer here
|
|
3516
3173
|
// is a lost precision gain on findings the run already paid for.
|
|
3517
3174
|
const CONSOLIDATE_TIMEOUT_MS = 10 * 60 * 1000;
|
|
3518
3175
|
const CONSOLIDATE_MAX_ATTEMPTS = 2;
|
|
@@ -3632,9 +3289,9 @@ async function runConsolidation(
|
|
|
3632
3289
|
name,
|
|
3633
3290
|
model: input.consolidate.model ?? input.model ?? "unresolved",
|
|
3634
3291
|
effort: defaultEffortForRole("other"),
|
|
3635
|
-
// FORCED to empty, never read from the prompt's frontmatter
|
|
3636
|
-
//
|
|
3637
|
-
//
|
|
3292
|
+
// FORCED to empty, never read from the prompt's frontmatter: the stage is
|
|
3293
|
+
// measured as claims-only, and a guarantee a prompt edit can revoke is not
|
|
3294
|
+
// a guarantee.
|
|
3638
3295
|
tools: [],
|
|
3639
3296
|
systemPromptPath,
|
|
3640
3297
|
outPath,
|
|
@@ -3644,8 +3301,8 @@ async function runConsolidation(
|
|
|
3644
3301
|
const startedAt = Date.now();
|
|
3645
3302
|
|
|
3646
3303
|
const abandon = (reason: string, error?: unknown): void => {
|
|
3647
|
-
//
|
|
3648
|
-
//
|
|
3304
|
+
// Keep the attempt count and pointers of a session that really ran; only a
|
|
3305
|
+
// pre-attempt failure lands its reason on the step.
|
|
3649
3306
|
if (
|
|
3650
3307
|
(meta.attempts === undefined || meta.attempts < 1) &&
|
|
3651
3308
|
meta.failure === undefined
|
|
@@ -4019,9 +3676,8 @@ async function runMaterialityJudge(
|
|
|
4019
3676
|
systemPromptPath,
|
|
4020
3677
|
prompt,
|
|
4021
3678
|
// FORCED to empty, never read from the prompt's frontmatter, for the
|
|
4022
|
-
//
|
|
4023
|
-
//
|
|
4024
|
-
// guarantee.
|
|
3679
|
+
// consolidation pass's reason: the judge is measured without tools, and a
|
|
3680
|
+
// guarantee a prompt edit can revoke is not a guarantee.
|
|
4025
3681
|
tools: [],
|
|
4026
3682
|
mcpConfigPath,
|
|
4027
3683
|
// PINNED, never `input.model` or the frontmatter: the measured judge is
|
|
@@ -4083,7 +3739,6 @@ async function runMaterialityJudge(
|
|
|
4083
3739
|
duration_ms: Date.now() - startedAt,
|
|
4084
3740
|
tokens_in: usage.tokens_in,
|
|
4085
3741
|
tokens_out: usage.tokens_out,
|
|
4086
|
-
cost_usd_est: usage.cost_usd_est,
|
|
4087
3742
|
attempts,
|
|
4088
3743
|
status: record.failed > 0 ? "failed" : "ok",
|
|
4089
3744
|
...(record.failed > 0 ? firstFailureDetail(settled) : {}),
|
|
@@ -4399,7 +4054,6 @@ function collectUnresolvedSpend(state: RunState): UnresolvedSpend[] {
|
|
|
4399
4054
|
step: step.name,
|
|
4400
4055
|
bucketId: reservation.bucketId,
|
|
4401
4056
|
reservationId: reservation.reservationId,
|
|
4402
|
-
knownUsd: reservation.knownUsd,
|
|
4403
4057
|
});
|
|
4404
4058
|
}
|
|
4405
4059
|
}
|
|
@@ -4478,14 +4132,13 @@ async function writePipelinePlan(
|
|
|
4478
4132
|
...(state.boundaryNonce === undefined
|
|
4479
4133
|
? {}
|
|
4480
4134
|
: { boundary_nonce: state.boundaryNonce }),
|
|
4481
|
-
...(state.scout === undefined ? {} : { scout: state.scout }),
|
|
4482
4135
|
...(state.consolidation === undefined
|
|
4483
4136
|
? {}
|
|
4484
4137
|
: { consolidation: state.consolidation }),
|
|
4485
|
-
// Unconditional, unlike the
|
|
4138
|
+
// Unconditional, unlike the one above: every road to this write (the
|
|
4486
4139
|
// gotchas early return included) records whether the gate was asked for.
|
|
4487
4140
|
materiality: state.materiality ?? skippedMateriality(input, state),
|
|
4488
|
-
// `--shared-prefix` arm identity, read off the artifact
|
|
4141
|
+
// `--shared-prefix` arm identity, read off the artifact: a run
|
|
4489
4142
|
// whose hunters all sit in groups of one has no warm-up row to tell it
|
|
4490
4143
|
// from control. Absent means the control layout, byte for byte.
|
|
4491
4144
|
...(input.sharedPrefix === true ? { shared_prefix: true } : {}),
|
|
@@ -4673,7 +4326,6 @@ function perAgentEntry(result: StepResult): PerAgentUsage {
|
|
|
4673
4326
|
duration_ms: result.usage.wall_ms,
|
|
4674
4327
|
tokens_in: result.usage.tokens_in,
|
|
4675
4328
|
tokens_out: result.usage.tokens_out,
|
|
4676
|
-
cost_usd_est: result.usage.cost_usd_est,
|
|
4677
4329
|
attempts: result.attempts,
|
|
4678
4330
|
status: result.status,
|
|
4679
4331
|
...detailField(result.status === "ok" ? undefined : result),
|
|
@@ -4707,7 +4359,6 @@ function failedAgentEntry(): PerAgentUsage {
|
|
|
4707
4359
|
duration_ms: 0,
|
|
4708
4360
|
tokens_in: 0,
|
|
4709
4361
|
tokens_out: 0,
|
|
4710
|
-
cost_usd_est: 0,
|
|
4711
4362
|
attempts: 0,
|
|
4712
4363
|
status: "failed",
|
|
4713
4364
|
};
|
|
@@ -4784,7 +4435,7 @@ function stepMeta(spec: StepSpec): StepMeta {
|
|
|
4784
4435
|
tools: spec.tools,
|
|
4785
4436
|
systemPromptPath: spec.systemPromptPath,
|
|
4786
4437
|
outPath: spec.outPath,
|
|
4787
|
-
// Pessimistic at construction
|
|
4438
|
+
// Pessimistic at construction: a step
|
|
4788
4439
|
// is pushed into the plan BEFORE it is spawned, and the ceiling can
|
|
4789
4440
|
// snapshot at any instant after that. Only a settled step overwrites this.
|
|
4790
4441
|
status: "unsettled",
|
|
@@ -4906,6 +4557,17 @@ function recordSettlement(
|
|
|
4906
4557
|
}
|
|
4907
4558
|
}
|
|
4908
4559
|
if (tracePaths.length > 0) meta.tracePaths = tracePaths;
|
|
4560
|
+
const usageFacts: AttemptUsageFacts[] = [];
|
|
4561
|
+
for (let attempt = 1; attempt <= result.attempts; attempt++) {
|
|
4562
|
+
const file = usageFactsPath(meta.outPath, meta.name, attempt);
|
|
4563
|
+
if (!existsSync(file)) continue;
|
|
4564
|
+
try {
|
|
4565
|
+
usageFacts.push(JSON.parse(readFileSync(file, "utf8")));
|
|
4566
|
+
} catch {
|
|
4567
|
+
// The harness writes it atomically; an unreadable one names no requests.
|
|
4568
|
+
}
|
|
4569
|
+
}
|
|
4570
|
+
if (usageFacts.length > 0) meta.usage_facts = usageFacts;
|
|
4909
4571
|
}
|
|
4910
4572
|
|
|
4911
4573
|
// A step that reached a terminal verdict WITHOUT a StepResult: the spec never
|