crysk 0.4.0 → 0.5.1
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.
- package/README.md +26 -17
- package/config/models/anthropic.json +1 -1
- package/dist/SKILL-k4k0zn1w.md +143 -0
- package/dist/SKILL-vhb4c3vv.md +553 -0
- package/dist/SKILL-xpf6kmdw.md +74 -0
- package/dist/ci-admission-0ken6ehx.md +128 -0
- package/dist/cli.js +3056 -1623
- package/dist/index.js +1744 -938
- package/dist/opencode-ci-ac5zksdn.md +106 -0
- package/dist/workflow-4pkyck01.yml +160 -0
- package/package.json +1 -1
- package/prompts/default/PROVENANCE.md +1 -0
- package/skills/crysk/SKILL.md +16 -16
- package/skills/crysk-ci-setup/SKILL.md +41 -38
- package/skills/crysk-ci-setup/assets/workflow.yml +31 -22
- package/skills/crysk-ci-setup/references/ci-admission.md +1 -1
- package/skills/crysk-ci-setup/references/opencode-ci.md +53 -88
- package/skills/crysk-loop/SKILL.md +1 -1
- package/src/asset-manifest.ts +0 -4
- package/src/assets.ts +1 -1
- package/src/ci/admission-ledger.ts +4 -0
- package/src/ci/classify-error.ts +21 -24
- package/src/ci/cli-record.ts +53 -1
- package/src/ci/crysk-secrets.ts +409 -0
- package/src/ci/force-comment.ts +14 -1
- package/src/ci/gates.ts +86 -57
- package/src/ci/outcome.ts +421 -330
- package/src/ci/platform-config.ts +2 -2
- package/src/ci/platform-estimate.ts +111 -0
- package/src/ci/platform-report.ts +274 -41
- package/src/ci/platform-trace.ts +581 -0
- package/src/ci/platform.ts +58 -12
- package/src/ci/publish.ts +19 -6
- package/src/ci/reporter.ts +14 -4
- package/src/ci/review-levels.ts +8 -7
- package/src/ci/run.ts +88 -0
- package/src/ci/setup.ts +485 -118
- package/src/ci/trace-bundle.ts +1486 -0
- package/src/ci/trace-transcripts.ts +480 -0
- package/src/ci/trigger.ts +176 -33
- package/src/cli.ts +11 -3
- package/src/commands/ci-setup-questions.ts +3 -3
- package/src/commands/ci-setup.ts +152 -6
- package/src/commands/config.ts +3 -0
- package/src/commands/doctor.ts +15 -6
- package/src/config/config.ts +16 -3
- package/src/diversity/clustering.ts +0 -1
- package/src/diversity/identity.ts +4 -5
- package/src/diversity/pipeline-integration.ts +1 -9
- package/src/doctor.ts +3 -3
- package/src/execution/attempt-evidence.ts +10 -5
- package/src/execution/claude-session-trace.ts +480 -0
- package/src/execution/contracts.ts +21 -4
- package/src/execution/harness.ts +105 -8
- package/src/execution/model-fallback.ts +239 -0
- package/src/execution/usage-normalized.ts +6 -4
- package/src/home-preflight.ts +1 -1
- package/src/index.ts +1 -1
- package/src/model/catalog.ts +13 -37
- package/src/model/free-discovery.ts +117 -8
- package/src/model/model-name.ts +404 -0
- package/src/model/provider-capabilities.ts +22 -7
- package/src/model/routing.ts +346 -467
- package/src/pr/admission.ts +18 -0
- package/src/pr/ci-admission-gate.ts +2 -0
- package/src/pr/discovery.ts +18 -0
- package/src/pr/plan.ts +55 -5
- package/src/pr/preflight.ts +1 -1
- package/src/pr/publish-outcome.ts +1 -1
- package/src/pr/review-pr.ts +26 -1
- package/src/pr/target.ts +63 -11
- package/src/production-runtime.ts +760 -331
- package/src/review/pipeline.ts +277 -101
- package/src/review/preflight.ts +201 -294
- package/src/review/report.ts +30 -26
- package/src/review/review.ts +20 -3
- package/src/review/route-preflight.ts +79 -11
- package/src/review/run.ts +44 -6
- package/src/review/scout.ts +1 -1
- package/src/review/spec.ts +3 -3
- package/src/review/step-runner.ts +70 -1
- package/src/runner-authority.ts +127 -9
- package/src/security/binding-policy.ts +1 -2
- package/src/security/claude-credential.ts +355 -0
- package/src/security/credential-broker.ts +56 -40
- package/src/security/evidence-redaction.ts +179 -4
- package/src/security/opencode-secret-manifest.ts +114 -0
- package/src/security/redact.ts +45 -0
- package/src/system-tools.ts +48 -21
- package/src/transport-registry.ts +12 -2
- package/src/transports/claude-code-cli.ts +21 -2
- package/src/transports/opencode-sdk.ts +11 -9
- package/src/transports/opencode-server.ts +7 -10
- package/src/ui/config-edit.ts +9 -20
- package/src/ui/config.ts +32 -27
- package/src/ui/plan.ts +84 -50
- package/src/updater.ts +10 -7
- package/prompts/default/deep-review-lifecycle.md +0 -362
package/src/review/pipeline.ts
CHANGED
|
@@ -11,23 +11,21 @@ import {
|
|
|
11
11
|
type MaterialityVerdict,
|
|
12
12
|
parseJudgeVerdict,
|
|
13
13
|
} from "#compare/materiality";
|
|
14
|
-
import {
|
|
15
|
-
|
|
16
|
-
type Effort,
|
|
17
|
-
type StepRole,
|
|
18
|
-
} from "#model/catalog";
|
|
14
|
+
import { defaultEffortForRole, type Effort, isEffort } from "#model/catalog";
|
|
15
|
+
import type { ModelsConfig } from "#model/model-name";
|
|
19
16
|
import {
|
|
20
17
|
agentStepKey,
|
|
21
18
|
buildResolvedRoutePlan,
|
|
19
|
+
claudeCliModel,
|
|
20
|
+
InvalidModelError,
|
|
22
21
|
isMeasuredJudgeRoute,
|
|
23
22
|
MATERIALITY_JUDGE_MODEL,
|
|
24
23
|
MATERIALITY_STEP_KEY,
|
|
25
24
|
type ResolvedModelRoute,
|
|
26
25
|
type ResolvedRoutePlan,
|
|
27
|
-
type
|
|
26
|
+
type ResolvedStepRoute,
|
|
28
27
|
type RunnerBackend,
|
|
29
28
|
resolveMaterialityJudgeRoute,
|
|
30
|
-
spawnModelForClaudeCli,
|
|
31
29
|
} from "#model/routing";
|
|
32
30
|
import {
|
|
33
31
|
composeAdjudicationPrompt,
|
|
@@ -140,14 +138,18 @@ import {
|
|
|
140
138
|
recordDiversityHunterResult,
|
|
141
139
|
} from "../diversity/pipeline-integration";
|
|
142
140
|
import { writeJsonAtomically } from "../execution/atomic-write";
|
|
143
|
-
import
|
|
144
|
-
|
|
145
|
-
|
|
141
|
+
import { claudeSessionTracePath } from "../execution/claude-session-trace";
|
|
142
|
+
import {
|
|
143
|
+
effortApplied,
|
|
144
|
+
type FailureCredential,
|
|
145
|
+
type TransportFailureCause,
|
|
146
146
|
} from "../execution/contracts";
|
|
147
|
+
import type { FailureCause } from "../execution/failure-policy";
|
|
147
148
|
// One predicate, owned beside the appends it recognizes. Already in this
|
|
148
149
|
// module's graph through step-runner's value import of the harness, so it
|
|
149
150
|
// adds no weight (the atomic-write WHY above is about a NEW dependency).
|
|
150
151
|
import { isHarnessStderrAnnotation } from "../execution/harness";
|
|
152
|
+
import { ranRoute } from "../execution/model-fallback";
|
|
151
153
|
import {
|
|
152
154
|
DEFAULT_CANCELLATION_DEADLINE_MS,
|
|
153
155
|
HARNESS_GRACE_MARGIN_MS,
|
|
@@ -294,9 +296,9 @@ export interface PipelineInput {
|
|
|
294
296
|
// demotes the immaterial ones to advisory. ABSENT means no judge step and
|
|
295
297
|
// findings exactly as before (the scout's rule); pipeline.json still
|
|
296
298
|
// carries a `materiality` record (`enabled: false`, `status: "skipped"`).
|
|
297
|
-
// No `model` seat on purpose: the judge
|
|
298
|
-
// measured on (
|
|
299
|
-
// frontmatter edit may move it.
|
|
299
|
+
// No `model` seat on purpose: the judge runs only on a model it was
|
|
300
|
+
// measured on (sonnet by default, opus through `models.judge`), and neither
|
|
301
|
+
// `--model`, `models.default` nor a frontmatter edit may move it.
|
|
300
302
|
materiality?: { promptPath: string };
|
|
301
303
|
// The PR title the materiality judge reads (S9): PrTarget.title in PR mode,
|
|
302
304
|
// the reviewed head's commit subject in local mode. Author-controlled, so
|
|
@@ -352,9 +354,11 @@ export interface PipelineInput {
|
|
|
352
354
|
// so callers that pass nothing see byte-identical step names, per_agent
|
|
353
355
|
// keys, and parity semantics.
|
|
354
356
|
spec?: ReviewSpec;
|
|
355
|
-
// D2 PR3: Optional pre-resolved model route plan or
|
|
357
|
+
// D2 PR3: Optional pre-resolved model route plan, or the role -> model
|
|
358
|
+
// lists to build one from (S1). Absent `models` falls back to the merged
|
|
359
|
+
// config's, as the routing config used to.
|
|
356
360
|
routePlan?: ResolvedRoutePlan;
|
|
357
|
-
|
|
361
|
+
models?: ModelsConfig;
|
|
358
362
|
// Item 7: skip hunter (and scout) fan-out. A re-review whose restricted
|
|
359
363
|
// delta is empty still classifies and verifies (C6); spawning hunters on
|
|
360
364
|
// an empty patch would bill a first-review for nothing. Absent = run
|
|
@@ -1090,6 +1094,10 @@ interface StepMeta {
|
|
|
1090
1094
|
// when the step carries no route (the legacy Claude-only shape), where the
|
|
1091
1095
|
// backend is not recorded and guessing it would be fabrication.
|
|
1092
1096
|
effort_applied?: boolean;
|
|
1097
|
+
// The OpenCode variant the step's backend was sent: its `models` entry's
|
|
1098
|
+
// effort (S7). Absent when none was sent (D1), and on every Claude step,
|
|
1099
|
+
// whose level is `effort` above.
|
|
1100
|
+
variant?: string;
|
|
1093
1101
|
tools: string[];
|
|
1094
1102
|
systemPromptPath: string;
|
|
1095
1103
|
outPath: string;
|
|
@@ -1131,6 +1139,10 @@ interface StepMeta {
|
|
|
1131
1139
|
attemptLogPath?: string;
|
|
1132
1140
|
attemptEvidencePath?: string;
|
|
1133
1141
|
settlementReceiptPath?: string;
|
|
1142
|
+
// Claude CLI session transcripts for every attempt whose copy exists,
|
|
1143
|
+
// relative to the run dir. Absent when none were copied, so a failed
|
|
1144
|
+
// capture names no file and a pipeline.json from before traces still parses.
|
|
1145
|
+
tracePaths?: string[];
|
|
1134
1146
|
// D1-08 PR2 (§8): the step's normalized usage, attached by recordSettlement
|
|
1135
1147
|
// — the single existing usage join site (D1-10c's own comment: a second
|
|
1136
1148
|
// pass over `state.steps` "would be a parallel mechanism able to disagree
|
|
@@ -1148,8 +1160,26 @@ interface StepMeta {
|
|
|
1148
1160
|
// not become a second, shorter copy of them. Redacted — a transport error
|
|
1149
1161
|
// can carry a token — and capped.
|
|
1150
1162
|
failure?: string;
|
|
1151
|
-
// D2 PR3: route attached at step construction
|
|
1163
|
+
// D2 PR3: route attached at step construction. On a step whose role listed
|
|
1164
|
+
// several models (S4), overwritten when it settles with the route of the
|
|
1165
|
+
// model that ran, like `model` and `effort_applied`.
|
|
1152
1166
|
route?: ResolvedModelRoute;
|
|
1167
|
+
// S4: every model the step tried, in order, with its own attempts, cause
|
|
1168
|
+
// and usage (models can bill differently, so only these keep each one's
|
|
1169
|
+
// billing). Absent on a step with one model.
|
|
1170
|
+
model_attempts?: Array<{
|
|
1171
|
+
model: string;
|
|
1172
|
+
route: ResolvedModelRoute;
|
|
1173
|
+
status: "ok" | "failed" | "skipped";
|
|
1174
|
+
attempts: number;
|
|
1175
|
+
// S7: the effort this model was asked to run at, and what its backend
|
|
1176
|
+
// was sent (as on the step itself).
|
|
1177
|
+
effort: Effort;
|
|
1178
|
+
effort_applied: boolean;
|
|
1179
|
+
variant?: string;
|
|
1180
|
+
cause?: FailureCause;
|
|
1181
|
+
usage_v2?: NormalizedUsage;
|
|
1182
|
+
}>;
|
|
1153
1183
|
}
|
|
1154
1184
|
|
|
1155
1185
|
// pipeline.json's `scout` key (§3.9). Written on EVERY run, including one
|
|
@@ -1548,8 +1578,7 @@ async function execute(
|
|
|
1548
1578
|
// it is settled long before a survivor exists to tier.
|
|
1549
1579
|
state.refuterConfigured = reviewSpec.agents.some((a) => a.role === "refuter");
|
|
1550
1580
|
|
|
1551
|
-
const
|
|
1552
|
-
input.routingConfig ?? input.config?.effective?.routing;
|
|
1581
|
+
const effectiveModels = input.models ?? input.config?.effective?.models;
|
|
1553
1582
|
|
|
1554
1583
|
const frontmatterByKey = new Map<string, string | undefined>();
|
|
1555
1584
|
for (const agent of reviewSpec.agents) {
|
|
@@ -1566,7 +1595,6 @@ async function execute(
|
|
|
1566
1595
|
let diversityCtx = prepareDiversityExecution({
|
|
1567
1596
|
reviewSpec,
|
|
1568
1597
|
cliModel: input.model,
|
|
1569
|
-
routingConfig: effectiveRoutingConfig,
|
|
1570
1598
|
frontmatterModel: (agentKey) => frontmatterByKey.get(agentKey),
|
|
1571
1599
|
target: input.diversityTarget,
|
|
1572
1600
|
runtimeTarget: reviewSpec.multiModelDiversity?.enabled
|
|
@@ -1617,7 +1645,7 @@ async function execute(
|
|
|
1617
1645
|
routePlan = buildResolvedRoutePlan({
|
|
1618
1646
|
agents: diversityCtx.routeAgents,
|
|
1619
1647
|
cliModel: input.model,
|
|
1620
|
-
|
|
1648
|
+
models: effectiveModels,
|
|
1621
1649
|
frontmatterModel: (agentKey) => frontmatterByKey.get(agentKey),
|
|
1622
1650
|
...(input.summarizer === undefined
|
|
1623
1651
|
? {}
|
|
@@ -1650,9 +1678,13 @@ async function execute(
|
|
|
1650
1678
|
: { materiality: true as const }),
|
|
1651
1679
|
});
|
|
1652
1680
|
} catch (error) {
|
|
1653
|
-
// Spec: legacy Claude-only runs without
|
|
1654
|
-
//
|
|
1655
|
-
|
|
1681
|
+
// Spec: legacy Claude-only runs without `models` keep today's direct
|
|
1682
|
+
// model passthrough (a raw Claude id the CLI resolves itself); D2 route
|
|
1683
|
+
// dimensions are optional then. A model written in the new form never
|
|
1684
|
+
// falls back: its error says how to write it.
|
|
1685
|
+
if (effectiveModels !== undefined || error instanceof InvalidModelError) {
|
|
1686
|
+
throw error;
|
|
1687
|
+
}
|
|
1656
1688
|
routePlan = undefined;
|
|
1657
1689
|
}
|
|
1658
1690
|
}
|
|
@@ -1830,14 +1862,17 @@ async function execute(
|
|
|
1830
1862
|
tools: agent.tools,
|
|
1831
1863
|
mcpConfigPath: input.mcpConfigPath,
|
|
1832
1864
|
model: resolveModel(input, hunter.model, agent.model, hunter.file),
|
|
1833
|
-
|
|
1865
|
+
...stepEffortFields(
|
|
1866
|
+
routePlan,
|
|
1867
|
+
name,
|
|
1868
|
+
agent.effort,
|
|
1869
|
+
defaultEffortForRole("hunter"),
|
|
1870
|
+
),
|
|
1834
1871
|
cwd: input.worktree,
|
|
1835
1872
|
outPath: path.join(stepsDir, `${name}.draft.json`),
|
|
1836
1873
|
timeoutMs: stepTimeoutMs,
|
|
1837
1874
|
maxAttempts: DEFAULT_STEP_MAX_ATTEMPTS,
|
|
1838
|
-
...(
|
|
1839
|
-
? {}
|
|
1840
|
-
: { route: routeForStepKey(routePlan, name) }),
|
|
1875
|
+
...stepRouteFields(routePlan, name),
|
|
1841
1876
|
parse: (finalText) => {
|
|
1842
1877
|
const extracted = extractJsonObject(finalText);
|
|
1843
1878
|
if (extracted === undefined) {
|
|
@@ -1896,7 +1931,13 @@ async function execute(
|
|
|
1896
1931
|
agent.model,
|
|
1897
1932
|
input.summarizer.promptPath,
|
|
1898
1933
|
);
|
|
1899
|
-
|
|
1934
|
+
const summarizerEffort = stepEffortFields(
|
|
1935
|
+
routePlan,
|
|
1936
|
+
name,
|
|
1937
|
+
agent.effort,
|
|
1938
|
+
defaultEffortForRole("summarizer"),
|
|
1939
|
+
);
|
|
1940
|
+
summarizerMeta.effort = summarizerEffort.effort;
|
|
1900
1941
|
summarizerMeta.tools = agent.tools;
|
|
1901
1942
|
summarizerSpec = {
|
|
1902
1943
|
name,
|
|
@@ -1905,12 +1946,10 @@ async function execute(
|
|
|
1905
1946
|
tools: agent.tools,
|
|
1906
1947
|
mcpConfigPath: input.mcpConfigPath,
|
|
1907
1948
|
model: summarizerMeta.model,
|
|
1908
|
-
|
|
1949
|
+
...summarizerEffort,
|
|
1909
1950
|
cwd: input.worktree,
|
|
1910
1951
|
outPath,
|
|
1911
|
-
...(
|
|
1912
|
-
? {}
|
|
1913
|
-
: { route: routeForStepKey(routePlan, name) }),
|
|
1952
|
+
...stepRouteFields(routePlan, name),
|
|
1914
1953
|
// The summary is a cosmetic barrier ahead of dedupe/refutation. It
|
|
1915
1954
|
// must not inherit the 30-minute hunter watchdog or its retry budget.
|
|
1916
1955
|
timeoutMs: 5 * 60 * 1000,
|
|
@@ -2473,7 +2512,12 @@ async function runRefuter(
|
|
|
2473
2512
|
agent.model,
|
|
2474
2513
|
options.agent.file,
|
|
2475
2514
|
);
|
|
2476
|
-
const
|
|
2515
|
+
const effortFields = stepEffortFields(
|
|
2516
|
+
state.routePlan,
|
|
2517
|
+
agentStepKey(options.agent),
|
|
2518
|
+
agent.effort,
|
|
2519
|
+
defaultEffortForRole("refuter"),
|
|
2520
|
+
);
|
|
2477
2521
|
// A finding's content is composed LONG after the run's nonce was committed
|
|
2478
2522
|
// — the hunters wrote its `claim` and `proof_refs` from the patch — so it is
|
|
2479
2523
|
// the one block `selectBoundaryNonce` could not be drawn against. Guarded
|
|
@@ -2516,7 +2560,7 @@ async function runRefuter(
|
|
|
2516
2560
|
tools: agent.tools,
|
|
2517
2561
|
mcpConfigPath: input.mcpConfigPath,
|
|
2518
2562
|
model,
|
|
2519
|
-
|
|
2563
|
+
...effortFields,
|
|
2520
2564
|
cwd: input.worktree,
|
|
2521
2565
|
outPath: path.join(
|
|
2522
2566
|
options.stepsDir,
|
|
@@ -2524,15 +2568,7 @@ async function runRefuter(
|
|
|
2524
2568
|
),
|
|
2525
2569
|
timeoutMs: options.stepTimeoutMs,
|
|
2526
2570
|
maxAttempts: DEFAULT_STEP_MAX_ATTEMPTS,
|
|
2527
|
-
...(
|
|
2528
|
-
undefined
|
|
2529
|
-
? {}
|
|
2530
|
-
: {
|
|
2531
|
-
route: routeForStepKey(
|
|
2532
|
-
state.routePlan,
|
|
2533
|
-
agentStepKey(options.agent),
|
|
2534
|
-
),
|
|
2535
|
-
}),
|
|
2571
|
+
...stepRouteFields(state.routePlan, agentStepKey(options.agent)),
|
|
2536
2572
|
parse: (finalText) => {
|
|
2537
2573
|
const extracted = extractJsonObject(finalText);
|
|
2538
2574
|
if (extracted === undefined) {
|
|
@@ -2707,7 +2743,12 @@ async function runVerify(
|
|
|
2707
2743
|
// The verifier re-reads the REFUTER's prompt file, so a frontmatter `effort:`
|
|
2708
2744
|
// there applies to it as well; `verifier` is only the fallback for a prompt
|
|
2709
2745
|
// that omits the key.
|
|
2710
|
-
const
|
|
2746
|
+
const effortFields = stepEffortFields(
|
|
2747
|
+
state.routePlan,
|
|
2748
|
+
agentStepKey(options.agent),
|
|
2749
|
+
agent.effort,
|
|
2750
|
+
defaultEffortForRole("verifier"),
|
|
2751
|
+
);
|
|
2711
2752
|
const forged: VerifySubject[] = [];
|
|
2712
2753
|
const specs: Array<{
|
|
2713
2754
|
subject: VerifySubject;
|
|
@@ -2729,7 +2770,7 @@ async function runVerify(
|
|
|
2729
2770
|
tools: agent.tools,
|
|
2730
2771
|
mcpConfigPath: input.mcpConfigPath,
|
|
2731
2772
|
model,
|
|
2732
|
-
|
|
2773
|
+
...effortFields,
|
|
2733
2774
|
cwd: input.worktree,
|
|
2734
2775
|
outPath: path.join(dir, "result.json"),
|
|
2735
2776
|
timeoutMs: options.stepTimeoutMs,
|
|
@@ -2738,15 +2779,7 @@ async function runVerify(
|
|
|
2738
2779
|
// admitted route, as the refuter and adjudicator legs do. Left unset, a
|
|
2739
2780
|
// routed run had no binding for these steps and every one failed
|
|
2740
2781
|
// "No admitted binding for step route" before reaching a provider.
|
|
2741
|
-
...(
|
|
2742
|
-
undefined
|
|
2743
|
-
? {}
|
|
2744
|
-
: {
|
|
2745
|
-
route: routeForStepKey(
|
|
2746
|
-
state.routePlan,
|
|
2747
|
-
agentStepKey(options.agent),
|
|
2748
|
-
),
|
|
2749
|
-
}),
|
|
2782
|
+
...stepRouteFields(state.routePlan, agentStepKey(options.agent)),
|
|
2750
2783
|
parse: (finalText) => {
|
|
2751
2784
|
const extracted = extractJsonObject(finalText);
|
|
2752
2785
|
if (extracted === undefined) {
|
|
@@ -2906,7 +2939,12 @@ async function runAdjudicate(
|
|
|
2906
2939
|
agent.model,
|
|
2907
2940
|
options.agent.file,
|
|
2908
2941
|
);
|
|
2909
|
-
const
|
|
2942
|
+
const effortFields = stepEffortFields(
|
|
2943
|
+
state.routePlan,
|
|
2944
|
+
agentStepKey(options.agent),
|
|
2945
|
+
agent.effort,
|
|
2946
|
+
defaultEffortForRole("refuter"),
|
|
2947
|
+
);
|
|
2910
2948
|
const specs: Array<{
|
|
2911
2949
|
index: number;
|
|
2912
2950
|
subject: AdjudicationSubject;
|
|
@@ -2927,20 +2965,12 @@ async function runAdjudicate(
|
|
|
2927
2965
|
tools: agent.tools,
|
|
2928
2966
|
mcpConfigPath: input.mcpConfigPath,
|
|
2929
2967
|
model,
|
|
2930
|
-
|
|
2968
|
+
...effortFields,
|
|
2931
2969
|
cwd: input.worktree,
|
|
2932
2970
|
outPath: path.join(dir, "result.json"),
|
|
2933
2971
|
timeoutMs: options.stepTimeoutMs,
|
|
2934
2972
|
maxAttempts: DEFAULT_STEP_MAX_ATTEMPTS,
|
|
2935
|
-
...(
|
|
2936
|
-
undefined
|
|
2937
|
-
? {}
|
|
2938
|
-
: {
|
|
2939
|
-
route: routeForStepKey(
|
|
2940
|
-
state.routePlan,
|
|
2941
|
-
agentStepKey(options.agent),
|
|
2942
|
-
),
|
|
2943
|
-
}),
|
|
2973
|
+
...stepRouteFields(state.routePlan, agentStepKey(options.agent)),
|
|
2944
2974
|
parse: parseAdjudicationResult,
|
|
2945
2975
|
onRetry: (info) => emit(deps, { kind: "step-retry", ...info }),
|
|
2946
2976
|
};
|
|
@@ -3143,7 +3173,13 @@ async function runScout(
|
|
|
3143
3173
|
input.scout.promptPath,
|
|
3144
3174
|
);
|
|
3145
3175
|
record.model = meta.model;
|
|
3146
|
-
|
|
3176
|
+
const scoutEffort = stepEffortFields(
|
|
3177
|
+
state.routePlan,
|
|
3178
|
+
name,
|
|
3179
|
+
agent.effort,
|
|
3180
|
+
defaultEffortForRole("scout"),
|
|
3181
|
+
);
|
|
3182
|
+
meta.effort = scoutEffort.effort;
|
|
3147
3183
|
// FULL sha256, not the 12 chars the probe prints: the probe's artifact
|
|
3148
3184
|
// stores the full digest too, so a pipeline.json and a scout-probe.json
|
|
3149
3185
|
// naming the same prompt say the same string.
|
|
@@ -3157,14 +3193,12 @@ async function runScout(
|
|
|
3157
3193
|
tools: meta.tools,
|
|
3158
3194
|
mcpConfigPath: input.mcpConfigPath,
|
|
3159
3195
|
model: meta.model,
|
|
3160
|
-
|
|
3196
|
+
...scoutEffort,
|
|
3161
3197
|
cwd: input.worktree,
|
|
3162
3198
|
outPath,
|
|
3163
3199
|
timeoutMs: SCOUT_TIMEOUT_MS,
|
|
3164
3200
|
maxAttempts: 1,
|
|
3165
|
-
...(
|
|
3166
|
-
? {}
|
|
3167
|
-
: { route: routeForStepKey(state.routePlan, name) }),
|
|
3201
|
+
...stepRouteFields(state.routePlan, name),
|
|
3168
3202
|
parse: (finalText) => {
|
|
3169
3203
|
const extracted = extractJsonObject(finalText);
|
|
3170
3204
|
if (extracted === undefined) {
|
|
@@ -3184,7 +3218,7 @@ async function runScout(
|
|
|
3184
3218
|
|
|
3185
3219
|
if (spec.route !== undefined) {
|
|
3186
3220
|
meta.route = spec.route;
|
|
3187
|
-
meta
|
|
3221
|
+
Object.assign(meta, routeEffortFields(spec.route));
|
|
3188
3222
|
const routedModel = effectiveExecutionModel(spec);
|
|
3189
3223
|
meta.model = routedModel;
|
|
3190
3224
|
record.model = routedModel;
|
|
@@ -3205,6 +3239,8 @@ async function runScout(
|
|
|
3205
3239
|
// BEFORE the verdict branches below, so every road that ends in `abandon`
|
|
3206
3240
|
// still carries the attempt count and pointers of the session that ran.
|
|
3207
3241
|
recordSettlement(meta, result, input.runDir);
|
|
3242
|
+
// The model that ran, when the scout fell back to another (S4).
|
|
3243
|
+
record.model = meta.model;
|
|
3208
3244
|
state.usageTotal = sumUsage(state.usageTotal, result.usage);
|
|
3209
3245
|
accumulateUsageV2(state, result);
|
|
3210
3246
|
if (result.status !== "ok") return abandon();
|
|
@@ -3413,7 +3449,13 @@ async function runConsolidation(
|
|
|
3413
3449
|
input.consolidate.promptPath,
|
|
3414
3450
|
);
|
|
3415
3451
|
record.model = meta.model;
|
|
3416
|
-
|
|
3452
|
+
const consolidateEffort = stepEffortFields(
|
|
3453
|
+
state.routePlan,
|
|
3454
|
+
name,
|
|
3455
|
+
agent.effort,
|
|
3456
|
+
DEFAULT_CONSOLIDATE_EFFORT,
|
|
3457
|
+
);
|
|
3458
|
+
meta.effort = consolidateEffort.effort;
|
|
3417
3459
|
record.prompt_sha256 = new Bun.CryptoHasher("sha256")
|
|
3418
3460
|
.update(raw)
|
|
3419
3461
|
.digest("hex");
|
|
@@ -3424,14 +3466,12 @@ async function runConsolidation(
|
|
|
3424
3466
|
tools: meta.tools,
|
|
3425
3467
|
mcpConfigPath: input.mcpConfigPath,
|
|
3426
3468
|
model: meta.model,
|
|
3427
|
-
|
|
3469
|
+
...consolidateEffort,
|
|
3428
3470
|
cwd: input.worktree,
|
|
3429
3471
|
outPath,
|
|
3430
3472
|
timeoutMs: CONSOLIDATE_TIMEOUT_MS,
|
|
3431
3473
|
maxAttempts: CONSOLIDATE_MAX_ATTEMPTS,
|
|
3432
|
-
...(
|
|
3433
|
-
? {}
|
|
3434
|
-
: { route: routeForStepKey(state.routePlan, name) }),
|
|
3474
|
+
...stepRouteFields(state.routePlan, name),
|
|
3435
3475
|
parse: (finalText) => {
|
|
3436
3476
|
const extracted = extractJsonObject(finalText);
|
|
3437
3477
|
if (extracted === undefined) {
|
|
@@ -3451,7 +3491,7 @@ async function runConsolidation(
|
|
|
3451
3491
|
|
|
3452
3492
|
if (spec.route !== undefined) {
|
|
3453
3493
|
meta.route = spec.route;
|
|
3454
|
-
meta
|
|
3494
|
+
Object.assign(meta, routeEffortFields(spec.route));
|
|
3455
3495
|
const routedModel = effectiveExecutionModel(spec);
|
|
3456
3496
|
meta.model = routedModel;
|
|
3457
3497
|
record.model = routedModel;
|
|
@@ -3469,6 +3509,8 @@ async function runConsolidation(
|
|
|
3469
3509
|
// tokens.
|
|
3470
3510
|
state.perAgent.consolidate = perAgentEntry(result);
|
|
3471
3511
|
recordSettlement(meta, result, input.runDir);
|
|
3512
|
+
// The model that ran, when consolidation fell back to another (S4).
|
|
3513
|
+
record.model = meta.model;
|
|
3472
3514
|
state.usageTotal = sumUsage(state.usageTotal, result.usage);
|
|
3473
3515
|
accumulateUsageV2(state, result);
|
|
3474
3516
|
if (result.status !== "ok") {
|
|
@@ -3614,12 +3656,14 @@ async function runMaterialityJudge(
|
|
|
3614
3656
|
record.reason =
|
|
3615
3657
|
route !== undefined ||
|
|
3616
3658
|
resolveMaterialityJudgeRoute(
|
|
3617
|
-
input.
|
|
3659
|
+
input.models ?? input.config?.effective?.models,
|
|
3618
3660
|
) === undefined
|
|
3619
3661
|
? "judge route is not the measured claude-code route"
|
|
3620
3662
|
: "route plan carries no judge step";
|
|
3621
3663
|
return;
|
|
3622
3664
|
}
|
|
3665
|
+
// `models.judge` may pick opus: the record names the model that runs.
|
|
3666
|
+
record.model = route.modelSnapshot;
|
|
3623
3667
|
}
|
|
3624
3668
|
if (judge.length === 0) return;
|
|
3625
3669
|
|
|
@@ -3680,7 +3724,16 @@ async function runMaterialityJudge(
|
|
|
3680
3724
|
input.prTitle !== undefined && input.prTitle.trim() !== ""
|
|
3681
3725
|
? input.prTitle
|
|
3682
3726
|
: MATERIALITY_TITLE_FALLBACK;
|
|
3683
|
-
|
|
3727
|
+
// models.judge takes no effort (D3), so this is the role default; it goes
|
|
3728
|
+
// through stepEffortFields anyway so no step builds its effort another way.
|
|
3729
|
+
// `.effort` only: the judge's list is chosen whole (judgeModelChoice) and
|
|
3730
|
+
// its step carries no fallbacks to pair efforts with.
|
|
3731
|
+
const { effort } = stepEffortFields(
|
|
3732
|
+
state.routePlan,
|
|
3733
|
+
MATERIALITY_STEP_KEY,
|
|
3734
|
+
undefined,
|
|
3735
|
+
defaultEffortForRole("other"),
|
|
3736
|
+
);
|
|
3684
3737
|
const specs: Array<{ id: string; spec: StepSpec; meta: StepMeta }> = [];
|
|
3685
3738
|
// The owner's file as posting policy, or undefined when it holds nothing of
|
|
3686
3739
|
// the owner's own: template text never reaches the judge (ownerPolicyText).
|
|
@@ -3725,9 +3778,10 @@ async function runMaterialityJudge(
|
|
|
3725
3778
|
tools: [],
|
|
3726
3779
|
mcpConfigPath,
|
|
3727
3780
|
// PINNED, never `input.model` or the frontmatter: the measured judge is
|
|
3728
|
-
// sonnet
|
|
3729
|
-
// deciding which
|
|
3730
|
-
|
|
3781
|
+
// sonnet or opus (`models.judge`, through the route), and a judge on
|
|
3782
|
+
// another model is a different, unmeasured judge deciding which
|
|
3783
|
+
// blockers post. A run with no route plan sends the measured default.
|
|
3784
|
+
model: route?.modelSnapshot ?? MATERIALITY_JUDGE_MODEL,
|
|
3731
3785
|
effort,
|
|
3732
3786
|
cwd: input.worktree,
|
|
3733
3787
|
outPath: path.join(dir, "result.json"),
|
|
@@ -4065,8 +4119,12 @@ function hunterFailureOf(
|
|
|
4065
4119
|
result: StepResult | undefined,
|
|
4066
4120
|
): HunterFailure {
|
|
4067
4121
|
const cause = result?.failureCause;
|
|
4068
|
-
|
|
4069
|
-
|
|
4122
|
+
// The model that ran, when the role listed several (S4): the provider and
|
|
4123
|
+
// model to name are the ones that failed last, not the first one listed.
|
|
4124
|
+
const route =
|
|
4125
|
+
(result === undefined ? undefined : ranRoute(result)) ?? spec.route;
|
|
4126
|
+
const provider = route?.provider.trim();
|
|
4127
|
+
const model = route?.modelSnapshot.trim();
|
|
4070
4128
|
const detail = failureDetailOf(result?.failureDetail);
|
|
4071
4129
|
return {
|
|
4072
4130
|
name: spec.name,
|
|
@@ -4297,14 +4355,56 @@ function resolveModel(
|
|
|
4297
4355
|
return model;
|
|
4298
4356
|
}
|
|
4299
4357
|
|
|
4300
|
-
//
|
|
4301
|
-
//
|
|
4302
|
-
//
|
|
4303
|
-
|
|
4358
|
+
// A step's effort AND each fallback's (S7), from the plan, as StepSpec fields.
|
|
4359
|
+
// One spread at every site that builds a step, beside stepRouteFields, so no
|
|
4360
|
+
// site can skip it or hand a fallback the first model's effort.
|
|
4361
|
+
//
|
|
4362
|
+
// Precedence, per model: its `models` entry's effort > the prompt's
|
|
4363
|
+
// frontmatter `effort:` > the role's default. There is no fourth seat on
|
|
4364
|
+
// purpose: nothing may fall through to the CLI's implicit default, which is
|
|
4365
|
+
// what #299 removed. `--model`, a spec model and a prompt alias carry no entry
|
|
4366
|
+
// effort (D4), so they keep frontmatter > role default.
|
|
4367
|
+
//
|
|
4368
|
+
// Only a Claude entry's effort lands in `effort`, the Claude scale the CLI's
|
|
4369
|
+
// `--effort` takes. An OpenCode entry's effort is a provider variant, already
|
|
4370
|
+
// on its route (`modelVariant`), and its step keeps the frontmatter or role
|
|
4371
|
+
// level as intent only, as before.
|
|
4372
|
+
//
|
|
4373
|
+
// `roleDefault` is a level, not a StepRole, because the consolidation pass has
|
|
4374
|
+
// its own default (DEFAULT_CONSOLIDATE_EFFORT) outside the role table.
|
|
4375
|
+
function stepEffortFields(
|
|
4376
|
+
routePlan: ResolvedRoutePlan | undefined,
|
|
4377
|
+
stepKey: string,
|
|
4304
4378
|
frontmatterEffort: Effort | undefined,
|
|
4305
|
-
|
|
4306
|
-
):
|
|
4307
|
-
|
|
4379
|
+
roleDefault: Effort,
|
|
4380
|
+
): Pick<StepSpec, "effort" | "fallbackEfforts"> {
|
|
4381
|
+
const base = frontmatterEffort ?? roleDefault;
|
|
4382
|
+
const step = routePlan?.steps.find((s) => s.stepKey === stepKey);
|
|
4383
|
+
if (step === undefined) return { effort: base };
|
|
4384
|
+
return {
|
|
4385
|
+
effort: entryEffort(step, base),
|
|
4386
|
+
...(step.fallbacks === undefined || step.fallbacks.length === 0
|
|
4387
|
+
? {}
|
|
4388
|
+
: {
|
|
4389
|
+
fallbackEfforts: step.fallbacks.map((fallback) =>
|
|
4390
|
+
entryEffort(fallback, base),
|
|
4391
|
+
),
|
|
4392
|
+
}),
|
|
4393
|
+
};
|
|
4394
|
+
}
|
|
4395
|
+
|
|
4396
|
+
function entryEffort(step: ResolvedStepRoute, base: Effort): Effort {
|
|
4397
|
+
if (step.route.backend !== "claude-code" || step.effort === undefined) {
|
|
4398
|
+
return base;
|
|
4399
|
+
}
|
|
4400
|
+
// parseModelList refused any other level, so this is an invariant, not a
|
|
4401
|
+
// reading: a level that reached here unvalidated must not run as another.
|
|
4402
|
+
if (!isEffort(step.effort)) {
|
|
4403
|
+
throw new Error(
|
|
4404
|
+
`${step.stepKey}: effort ${JSON.stringify(step.effort)} is not a Claude effort level`,
|
|
4405
|
+
);
|
|
4406
|
+
}
|
|
4407
|
+
return step.effort;
|
|
4308
4408
|
}
|
|
4309
4409
|
|
|
4310
4410
|
function perAgentEntry(result: StepResult): PerAgentUsage {
|
|
@@ -4377,17 +4477,42 @@ function routeForStepKey(
|
|
|
4377
4477
|
return routePlan?.steps.find((step) => step.stepKey === stepKey)?.route;
|
|
4378
4478
|
}
|
|
4379
4479
|
|
|
4480
|
+
// A step's route AND its fallbacks (S4), from the plan, as StepSpec fields.
|
|
4481
|
+
// One spread at every site that routes a step, so no site can hand a runner
|
|
4482
|
+
// the first model without the rest of the list.
|
|
4483
|
+
function stepRouteFields(
|
|
4484
|
+
routePlan: ResolvedRoutePlan | undefined,
|
|
4485
|
+
stepKey: string,
|
|
4486
|
+
): Pick<StepSpec, "route" | "fallbackRoutes"> {
|
|
4487
|
+
const step = routePlan?.steps.find((s) => s.stepKey === stepKey);
|
|
4488
|
+
if (step === undefined) return {};
|
|
4489
|
+
return {
|
|
4490
|
+
route: step.route,
|
|
4491
|
+
...(step.fallbacks === undefined || step.fallbacks.length === 0
|
|
4492
|
+
? {}
|
|
4493
|
+
: { fallbackRoutes: step.fallbacks.map((fallback) => fallback.route) }),
|
|
4494
|
+
};
|
|
4495
|
+
}
|
|
4496
|
+
|
|
4497
|
+
// What a route's backend is sent: the Claude CLI's `--model`, or the
|
|
4498
|
+
// provider's model id.
|
|
4499
|
+
function routeExecutionModel(
|
|
4500
|
+
route: ResolvedModelRoute,
|
|
4501
|
+
fallback: string,
|
|
4502
|
+
): string {
|
|
4503
|
+
if (route.backend === "claude-code") {
|
|
4504
|
+
return claudeCliModel(route);
|
|
4505
|
+
}
|
|
4506
|
+
return route.modelSnapshot || fallback;
|
|
4507
|
+
}
|
|
4508
|
+
|
|
4509
|
+
// What the step's backend is sent, recorded as the step's model in
|
|
4510
|
+
// pipeline.json: the route's model when there is a route, else the step's own.
|
|
4380
4511
|
function effectiveExecutionModel(spec: StepSpec): string {
|
|
4381
4512
|
if (spec.route === undefined) {
|
|
4382
4513
|
return spec.model;
|
|
4383
4514
|
}
|
|
4384
|
-
|
|
4385
|
-
return spawnModelForClaudeCli(spec.route, spec.model);
|
|
4386
|
-
}
|
|
4387
|
-
return (
|
|
4388
|
-
(spec.route as { model?: string } | undefined)?.model ??
|
|
4389
|
-
(spec.route.modelSnapshot || spec.model)
|
|
4390
|
-
);
|
|
4515
|
+
return routeExecutionModel(spec.route, spec.model);
|
|
4391
4516
|
}
|
|
4392
4517
|
|
|
4393
4518
|
function stepMeta(spec: StepSpec): StepMeta {
|
|
@@ -4395,9 +4520,7 @@ function stepMeta(spec: StepSpec): StepMeta {
|
|
|
4395
4520
|
name: spec.name,
|
|
4396
4521
|
model: effectiveExecutionModel(spec),
|
|
4397
4522
|
effort: spec.effort,
|
|
4398
|
-
...(spec.route === undefined
|
|
4399
|
-
? {}
|
|
4400
|
-
: { effort_applied: spec.route.backend === "claude-code" }),
|
|
4523
|
+
...(spec.route === undefined ? {} : routeEffortFields(spec.route)),
|
|
4401
4524
|
tools: spec.tools,
|
|
4402
4525
|
systemPromptPath: spec.systemPromptPath,
|
|
4403
4526
|
outPath: spec.outPath,
|
|
@@ -4409,6 +4532,22 @@ function stepMeta(spec: StepSpec): StepMeta {
|
|
|
4409
4532
|
};
|
|
4410
4533
|
}
|
|
4411
4534
|
|
|
4535
|
+
// What pipeline.json says a route's backend was sent for effort (S7): whether
|
|
4536
|
+
// it applied one (effortApplied) and, on OpenCode, the variant it was sent.
|
|
4537
|
+
// `variant` is absent when none was sent, so a run with no entry effort
|
|
4538
|
+
// records what it always recorded.
|
|
4539
|
+
function routeEffortFields(route: ResolvedModelRoute): {
|
|
4540
|
+
effort_applied: boolean;
|
|
4541
|
+
variant?: string;
|
|
4542
|
+
} {
|
|
4543
|
+
return {
|
|
4544
|
+
effort_applied: effortApplied(route),
|
|
4545
|
+
...(route.modelVariant === undefined
|
|
4546
|
+
? {}
|
|
4547
|
+
: { variant: route.modelVariant }),
|
|
4548
|
+
};
|
|
4549
|
+
}
|
|
4550
|
+
|
|
4412
4551
|
// D1-10c. Attaches one settled step's provenance from the SAME `StepResult`
|
|
4413
4552
|
// the usage join already reads, at the SAME join site — a second pass over the
|
|
4414
4553
|
// steps would be a parallel mechanism able to disagree with `per_agent`, and
|
|
@@ -4439,6 +4578,35 @@ function recordSettlement(
|
|
|
4439
4578
|
if (result.reservations !== undefined && result.reservations.length > 0) {
|
|
4440
4579
|
meta.reservations = result.reservations;
|
|
4441
4580
|
}
|
|
4581
|
+
// S4: a step whose role listed several models records the one that ran,
|
|
4582
|
+
// over the first model stamped before the spawn, and every model it tried.
|
|
4583
|
+
// The backend and the effort can change with the model (S7: each runs at
|
|
4584
|
+
// its own), so `effort`, `effort_applied` and `variant` follow the model
|
|
4585
|
+
// that ran. A step with one model carries no `modelAttempts` and records
|
|
4586
|
+
// what it always recorded.
|
|
4587
|
+
if (result.modelAttempts !== undefined) {
|
|
4588
|
+
const firstEffort = meta.effort;
|
|
4589
|
+
const ran = result.modelAttempts.findLast(
|
|
4590
|
+
(entry) => entry.status !== "skipped",
|
|
4591
|
+
);
|
|
4592
|
+
if (ran !== undefined) {
|
|
4593
|
+
meta.route = ran.route;
|
|
4594
|
+
meta.model = routeExecutionModel(ran.route, meta.model);
|
|
4595
|
+
meta.effort = ran.effort ?? firstEffort;
|
|
4596
|
+
delete meta.variant;
|
|
4597
|
+
Object.assign(meta, routeEffortFields(ran.route));
|
|
4598
|
+
}
|
|
4599
|
+
meta.model_attempts = result.modelAttempts.map((entry) => ({
|
|
4600
|
+
model: routeExecutionModel(entry.route, meta.model),
|
|
4601
|
+
route: entry.route,
|
|
4602
|
+
status: entry.status,
|
|
4603
|
+
attempts: entry.attempts,
|
|
4604
|
+
effort: entry.effort ?? firstEffort,
|
|
4605
|
+
...routeEffortFields(entry.route),
|
|
4606
|
+
...(entry.cause === undefined ? {} : { cause: entry.cause }),
|
|
4607
|
+
...(entry.usageV2 === undefined ? {} : { usage_v2: entry.usageV2 }),
|
|
4608
|
+
}));
|
|
4609
|
+
}
|
|
4442
4610
|
meta.status = result.status;
|
|
4443
4611
|
meta.attempts = result.attempts;
|
|
4444
4612
|
// Cancellation before the first admitted attempt, and every preflight
|
|
@@ -4470,6 +4638,14 @@ function recordSettlement(
|
|
|
4470
4638
|
runDir,
|
|
4471
4639
|
settlementReceiptPath(meta.outPath, meta.name, result.attempts),
|
|
4472
4640
|
);
|
|
4641
|
+
const tracePaths: string[] = [];
|
|
4642
|
+
for (let attempt = 1; attempt <= result.attempts; attempt++) {
|
|
4643
|
+
const absolute = claudeSessionTracePath(meta.outPath, meta.name, attempt);
|
|
4644
|
+
if (existsSync(absolute)) {
|
|
4645
|
+
tracePaths.push(path.relative(runDir, absolute));
|
|
4646
|
+
}
|
|
4647
|
+
}
|
|
4648
|
+
if (tracePaths.length > 0) meta.tracePaths = tracePaths;
|
|
4473
4649
|
}
|
|
4474
4650
|
|
|
4475
4651
|
// A step that reached a terminal verdict WITHOUT a StepResult: the spec never
|