principles-disciple 1.71.0 → 1.73.0
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/openclaw.plugin.json +10 -5
- package/package.json +17 -19
- package/scripts/acceptance-test.mjs +16 -73
- package/scripts/sync-plugin.mjs +382 -77
- package/src/commands/archive-impl.ts +2 -1
- package/src/commands/capabilities.ts +2 -2
- package/src/commands/context.ts +2 -2
- package/src/commands/disable-impl.ts +2 -1
- package/src/commands/evolution-status.ts +16 -16
- package/src/commands/export.ts +12 -67
- package/src/commands/pain.ts +91 -1
- package/src/commands/principle-rollback.ts +2 -1
- package/src/commands/promote-impl.ts +7 -43
- package/src/commands/rollback-impl.ts +2 -1
- package/src/commands/rollback.ts +2 -1
- package/src/commands/samples.ts +2 -1
- package/src/commands/thinking-os.ts +2 -1
- package/src/config/errors.ts +18 -2
- package/src/constants/diagnostician.ts +2 -2
- package/src/constants/tools.ts +2 -1
- package/src/core/__tests__/focus-history.test.ts +210 -0
- package/src/core/config.ts +1 -1
- package/src/core/confirm-first-gate.ts +255 -0
- package/src/core/correction-cue-learner.ts +2 -136
- package/src/core/correction-types.ts +16 -88
- package/src/core/dictionary.ts +19 -20
- package/src/core/empathy-keyword-matcher.ts +17 -289
- package/src/core/empathy-types.ts +18 -229
- package/src/core/event-log.ts +38 -132
- package/src/core/evolution-reducer.ts +21 -2
- package/src/core/evolution-types.ts +76 -464
- package/src/core/file-store.ts +80 -0
- package/src/core/focus-history.ts +228 -955
- package/src/core/local-worker-routing.ts +34 -314
- package/src/core/merge-gate-audit.ts +0 -195
- package/src/core/pain-diagnostic-gate.ts +154 -0
- package/src/core/pain-signal.ts +21 -138
- package/src/core/pain.ts +15 -88
- package/src/core/pd-task-reconciler.ts +26 -115
- package/src/core/pd-task-service.ts +9 -9
- package/src/core/pd-task-types.ts +23 -127
- package/src/core/principle-compiler/__tests__/compiler-replay-gate.test.ts +174 -0
- package/src/core/principle-compiler/code-validator.ts +15 -42
- package/src/core/principle-compiler/compiler.ts +100 -15
- package/src/core/principle-compiler/index.ts +5 -2
- package/src/core/principle-compiler/template-generator.ts +4 -104
- package/src/core/principle-injection.ts +10 -202
- package/src/core/principle-internalization/filesystem-lifecycle-datasource.ts +42 -0
- package/src/core/principle-internalization/lifecycle-read-model.ts +39 -242
- package/src/core/principle-internalization/principle-lifecycle-service.ts +12 -10
- package/src/core/principle-tree-ledger-adapter.ts +145 -0
- package/src/core/principle-tree-ledger.ts +8 -6
- package/src/core/reflection/reflection-context.ts +14 -109
- package/src/core/replay-engine.ts +8 -500
- package/src/core/rule-host-helpers.ts +5 -35
- package/src/core/rule-host-types.ts +10 -82
- package/src/core/rule-host.ts +6 -63
- package/src/core/runtime-v2-prompt-activation-reader.ts +231 -0
- package/src/core/session-tracker.ts +87 -101
- package/src/core/shadow-observation-registry.ts +19 -48
- package/src/core/trajectory.ts +3 -1
- package/src/core/workflow-funnel-loader.ts +62 -68
- package/src/core/workspace-context.ts +46 -0
- package/src/core/workspace-dir-service.ts +1 -1
- package/src/core/workspace-dir-validation.ts +18 -9
- package/src/hooks/AGENTS.md +1 -1
- package/src/hooks/gate-block-helper.ts +46 -44
- package/src/hooks/gate.ts +207 -7
- package/src/hooks/lifecycle.ts +30 -32
- package/src/hooks/llm.ts +60 -32
- package/src/hooks/pain.ts +297 -103
- package/src/hooks/prompt.ts +469 -339
- package/src/hooks/subagent.ts +2 -29
- package/src/i18n/commands.ts +2 -10
- package/src/index.ts +95 -85
- package/src/openclaw-sdk.ts +311 -0
- package/src/service/central-database.ts +8 -4
- package/src/service/evolution-queue-migration.ts +2 -1
- package/src/service/evolution-worker.ts +163 -1786
- package/src/service/internalization-trigger-adapter.ts +302 -0
- package/src/service/keyword-optimization-service.ts +4 -4
- package/src/service/monitoring-query-service.ts +1 -215
- package/src/service/queue-io.ts +60 -331
- package/src/service/runtime-summary-service.ts +115 -18
- package/src/service/subagent-workflow/index.ts +0 -41
- package/src/service/subagent-workflow/types.ts +9 -120
- package/src/service/subagent-workflow/workflow-store.ts +2 -119
- package/src/service/workflow-watchdog.ts +0 -43
- package/src/types/event-payload.ts +16 -74
- package/src/types/event-types.ts +39 -547
- package/src/types/hygiene-types.ts +7 -30
- package/src/types/principle-tree-schema.ts +20 -222
- package/src/types/queue.ts +15 -70
- package/src/types/runtime-summary.ts +5 -49
- package/src/utils/io.ts +10 -0
- package/src/utils/retry.ts +1 -1
- package/src/utils/shadow-fingerprint.ts +2 -2
- package/src/utils/workspace-resolver.ts +50 -0
- package/templates/langs/en/core/AGENTS.md +2 -2
- package/templates/langs/en/core/BOOT.md +1 -1
- package/templates/langs/en/core/HEARTBEAT.md +2 -2
- package/templates/langs/en/skills/ai-sprint-orchestration/references/agent-registry.json +1 -72
- package/templates/langs/en/skills/ai-sprint-orchestration/references/specs/bugfix-complex-template.json +6 -6
- package/templates/langs/en/skills/ai-sprint-orchestration/references/specs/feature-complex-template.json +6 -6
- package/templates/langs/en/skills/ai-sprint-orchestration/references/specs/workflow-validation-minimal-verify.json +2 -12
- package/templates/langs/en/skills/ai-sprint-orchestration/references/specs/workflow-validation-minimal.json +2 -12
- package/templates/langs/en/skills/ai-sprint-orchestration/runtime/.gitignore +2 -2
- package/templates/langs/en/skills/ai-sprint-orchestration/scripts/run.mjs +51 -15
- package/templates/langs/en/skills/evolve-task/SKILL.md +1 -1
- package/templates/langs/en/skills/pd-cli-operator/SKILL.md +67 -0
- package/templates/langs/en/skills/pd-diagnostician/SKILL.md +1 -1
- package/templates/langs/en/skills/pd-mentor/SKILL.md +1 -1
- package/templates/langs/en/skills/pd-pain-signal/SKILL.md +17 -39
- package/templates/langs/en/skills/pd-runtime-v2/SKILL.md +61 -0
- package/templates/langs/zh/core/AGENTS.md +2 -2
- package/templates/langs/zh/core/BOOT.md +1 -1
- package/templates/langs/zh/core/HEARTBEAT.md +2 -2
- package/templates/langs/zh/skills/ai-sprint-orchestration/references/agent-registry.json +1 -72
- package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/bugfix-complex-template.json +6 -6
- package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/feature-complex-template.json +6 -6
- package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/nocturnal-trinity-quality-enhancement.json +8 -8
- package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/workflow-validation-minimal-verify.json +2 -12
- package/templates/langs/zh/skills/ai-sprint-orchestration/references/specs/workflow-validation-minimal.json +2 -12
- package/templates/langs/zh/skills/ai-sprint-orchestration/runtime/.gitignore +2 -2
- package/templates/langs/zh/skills/ai-sprint-orchestration/scripts/run.mjs +51 -15
- package/templates/langs/zh/skills/ai-sprint-orchestration/test/run.test.mjs +21 -5
- package/templates/langs/zh/skills/evolve-task/SKILL.md +2 -2
- package/templates/langs/zh/skills/pd-cli-operator/SKILL.md +67 -0
- package/templates/langs/zh/skills/pd-diagnostician/SKILL.md +1 -1
- package/templates/langs/zh/skills/pd-mentor/SKILL.md +1 -1
- package/templates/langs/zh/skills/pd-pain-signal/SKILL.md +17 -38
- package/templates/langs/zh/skills/pd-runtime-v2/SKILL.md +61 -0
- package/tests/build-artifacts.test.ts +1 -3
- package/tests/commands/evolution-status.test.ts +0 -118
- package/tests/core/bootstrap-rules.test.ts +1 -1
- package/tests/core/config.test.ts +1 -1
- package/tests/core/event-log.test.ts +35 -0
- package/tests/core/evolution-engine.test.ts +610 -0
- package/tests/core/file-store.test.ts +102 -0
- package/tests/core/focus-history.test.ts +203 -11
- package/tests/core/merge-gate-audit.test.ts +2 -169
- package/tests/core/model-deployment-registry.test.ts +7 -1
- package/tests/core/model-training-registry.test.ts +19 -0
- package/tests/core/observability.test.ts +0 -1
- package/tests/core/pain-diagnostic-gate.test.ts +498 -0
- package/tests/core/pain.test.ts +0 -1
- package/tests/core/principle-internalization/deprecated-readiness.test.ts +2 -2
- package/tests/core/principle-internalization/lifecycle-metrics.test.ts +2 -2
- package/tests/core/principle-internalization/{internalization-routing-policy.test.ts → lifecycle-routing-policy.test.ts} +6 -6
- package/tests/core/principle-internalization/lineage-source-retired.test.ts +56 -0
- package/tests/core/principle-internalization/principle-lifecycle-service.test.ts +1 -23
- package/tests/core/principle-tree-ledger-adapter.test.ts +253 -0
- package/tests/core/reflection-context.test.ts +0 -14
- package/tests/core/replay-engine.test.ts +127 -215
- package/tests/core/rule-host-helpers.test.ts +2 -2
- package/tests/core/rule-implementation-runtime.test.ts +0 -27
- package/tests/core/workflow-funnel-loader.test.ts +162 -0
- package/tests/core/workspace-dir-validation.test.ts +8 -1
- package/tests/core-anti-growth.test.ts +192 -0
- package/tests/hook-workspace-nextaction-contract.test.ts +42 -0
- package/tests/hooks/confirm-first-gate.test.ts +333 -0
- package/tests/hooks/gate-auto-correct-shadow.test.ts +310 -0
- package/tests/hooks/gate-auto-correct.test.ts +665 -0
- package/tests/hooks/gate-rule-host-pipeline.test.ts +2 -1
- package/tests/hooks/pain.test.ts +269 -12
- package/tests/hooks/prompt-characterization.test.ts +500 -0
- package/tests/hooks/prompt-size-guard.test.ts +329 -0
- package/tests/hooks/runtime-v2-prompt-activation.test.ts +869 -0
- package/tests/index.test.ts +94 -1
- package/tests/integration/auto-entry-gate.test.ts +248 -0
- package/tests/integration/internalization-trigger-guard.test.ts +69 -0
- package/tests/integration/m8-legacy-paths.test.ts +63 -0
- package/tests/integration/runtime-v2-pain-guard.test.ts +125 -0
- package/tests/plugin-config-resolution-cutover.test.ts +359 -0
- package/tests/runtime-v2-discovery-guard.test.ts +154 -0
- package/tests/service/central-database.test.ts +457 -0
- package/tests/service/evolution-worker.correction-observer.test.ts +173 -0
- package/tests/service/evolution-worker.timeout.test.ts +11 -129
- package/tests/service/internalization-trigger-adapter.test.ts +251 -0
- package/tests/service/monitoring-query-service.test.ts +1 -47
- package/tests/service/queue-io.test.ts +1 -62
- package/tests/service/runtime-summary-service.test.ts +184 -3
- package/tests/service/workflow-watchdog.test.ts +0 -91
- package/tests/utils/file-lock.test.ts +5 -3
- package/tests/utils/session-key.test.ts +52 -0
- package/tests/utils/subagent-probe.test.ts +48 -1
- package/vitest.config.ts +4 -11
- package/.planning/codebase/ARCHITECTURE.md +0 -157
- package/.planning/codebase/CONCERNS.md +0 -145
- package/.planning/codebase/CONVENTIONS.md +0 -148
- package/.planning/codebase/INTEGRATIONS.md +0 -81
- package/.planning/codebase/STACK.md +0 -87
- package/.planning/codebase/STRUCTURE.md +0 -193
- package/.planning/codebase/TESTING.md +0 -243
- package/.planning/phases/01-basic-visualization/01-GAP-CLOSURE-VERIFICATION.md +0 -113
- package/docs/COMMAND_REFERENCE.md +0 -76
- package/docs/COMMAND_REFERENCE_EN.md +0 -79
- package/scripts/build-web.mjs +0 -46
- package/scripts/diagnose-nocturnal.mjs +0 -537
- package/scripts/seed-nocturnal-scenarios.mjs +0 -384
- package/src/commands/nocturnal-review.ts +0 -322
- package/src/commands/nocturnal-rollout.ts +0 -790
- package/src/commands/nocturnal-train.ts +0 -986
- package/src/commands/pd-reflect.ts +0 -88
- package/src/core/adaptive-thresholds.ts +0 -478
- package/src/core/diagnostician-task-store.ts +0 -192
- package/src/core/nocturnal-arbiter.ts +0 -715
- package/src/core/nocturnal-artifact-lineage.ts +0 -116
- package/src/core/nocturnal-artificer.ts +0 -257
- package/src/core/nocturnal-candidate-scoring.ts +0 -530
- package/src/core/nocturnal-compliance.ts +0 -1146
- package/src/core/nocturnal-dataset.ts +0 -763
- package/src/core/nocturnal-executability.ts +0 -428
- package/src/core/nocturnal-export.ts +0 -499
- package/src/core/nocturnal-paths.ts +0 -240
- package/src/core/nocturnal-reasoning-deriver.ts +0 -343
- package/src/core/nocturnal-rule-implementation-validator.ts +0 -246
- package/src/core/nocturnal-snapshot-contract.ts +0 -99
- package/src/core/nocturnal-trajectory-extractor.ts +0 -512
- package/src/core/nocturnal-trinity-types.ts +0 -218
- package/src/core/nocturnal-trinity.ts +0 -2680
- package/src/core/principle-internalization/deprecated-readiness.ts +0 -93
- package/src/core/principle-internalization/internalization-routing-policy.ts +0 -208
- package/src/core/principle-internalization/lifecycle-metrics.ts +0 -152
- package/src/http/principles-console-route.ts +0 -709
- package/src/service/central-health-service.ts +0 -49
- package/src/service/central-overview-service.ts +0 -138
- package/src/service/control-ui-query-service.ts +0 -900
- package/src/service/cooldown-strategy.ts +0 -97
- package/src/service/evolution-pain-context.ts +0 -79
- package/src/service/evolution-query-service.ts +0 -407
- package/src/service/health-query-service.ts +0 -1038
- package/src/service/nocturnal-config.ts +0 -214
- package/src/service/nocturnal-runtime.ts +0 -734
- package/src/service/nocturnal-service.ts +0 -1605
- package/src/service/nocturnal-target-selector.ts +0 -545
- package/src/service/sleep-cycle.ts +0 -157
- package/src/service/startup-reconciler.ts +0 -112
- package/src/service/subagent-workflow/correction-observer-types.ts +0 -82
- package/src/service/subagent-workflow/correction-observer-workflow-manager.ts +0 -250
- package/src/service/subagent-workflow/deep-reflect-workflow-manager.ts +0 -1
- package/src/service/subagent-workflow/dynamic-timeout.ts +0 -30
- package/src/service/subagent-workflow/empathy-observer-workflow-manager.ts +0 -268
- package/src/service/subagent-workflow/nocturnal-workflow-manager.ts +0 -795
- package/src/service/subagent-workflow/runtime-direct-driver.ts +0 -268
- package/src/service/subagent-workflow/workflow-manager-base.ts +0 -580
- package/src/tools/write-pain-flag.ts +0 -215
- package/tests/commands/nocturnal-review.test.ts +0 -448
- package/tests/commands/nocturnal-train.test.ts +0 -97
- package/tests/commands/pd-reflect.test.ts +0 -49
- package/tests/core/adaptive-thresholds.test.ts +0 -261
- package/tests/core/nocturnal-arbiter.test.ts +0 -559
- package/tests/core/nocturnal-artifact-lineage.test.ts +0 -53
- package/tests/core/nocturnal-artificer.test.ts +0 -241
- package/tests/core/nocturnal-candidate-scoring.test.ts +0 -532
- package/tests/core/nocturnal-compliance-p-principles.test.ts +0 -133
- package/tests/core/nocturnal-compliance.test.ts +0 -646
- package/tests/core/nocturnal-dataset.test.ts +0 -892
- package/tests/core/nocturnal-e2e.test.ts +0 -234
- package/tests/core/nocturnal-executability.test.ts +0 -357
- package/tests/core/nocturnal-export.test.ts +0 -517
- package/tests/core/nocturnal-reasoning-deriver.test.ts +0 -372
- package/tests/core/nocturnal-reviewed-subset-comparison.test.ts +0 -428
- package/tests/core/nocturnal-rule-implementation-validator.test.ts +0 -127
- package/tests/core/nocturnal-snapshot-contract.test.ts +0 -121
- package/tests/core/nocturnal-trajectory-extractor.test.ts +0 -634
- package/tests/core/nocturnal-trinity.test.ts +0 -2053
- package/tests/core/pain-auto-repair.test.ts +0 -96
- package/tests/core/pain-integration.test.ts +0 -510
- package/tests/fixtures/nocturnal-reviewed-subset.json +0 -183
- package/tests/http/principles-console-route.test.ts +0 -162
- package/tests/integration/chaos-resilience.test.ts +0 -348
- package/tests/integration/empathy-workflow-integration.test.ts +0 -626
- package/tests/integration/pain-diagnostician-loop.e2e.test.ts +0 -380
- package/tests/service/control-ui-query-service.test.ts +0 -121
- package/tests/service/cooldown-strategy.test.ts +0 -164
- package/tests/service/data-endpoints-regression.test.ts +0 -834
- package/tests/service/empathy-observer-workflow-manager.test.ts +0 -175
- package/tests/service/evolution-worker.nocturnal.test.ts +0 -601
- package/tests/service/nocturnal-runtime-hardening.test.ts +0 -118
- package/tests/service/nocturnal-runtime.test.ts +0 -473
- package/tests/service/nocturnal-service-code-candidate.test.ts +0 -330
- package/tests/service/nocturnal-target-selector.test.ts +0 -615
- package/tests/service/startup-reconciler.test.ts +0 -148
- package/tests/tools/write-pain-flag.test.ts +0 -358
- package/ui/src/App.tsx +0 -45
- package/ui/src/api.ts +0 -220
- package/ui/src/charts.tsx +0 -955
- package/ui/src/components/ErrorState.tsx +0 -6
- package/ui/src/components/Loading.tsx +0 -13
- package/ui/src/components/ProtectedRoute.tsx +0 -12
- package/ui/src/components/Shell.tsx +0 -91
- package/ui/src/components/WorkspaceConfig.tsx +0 -178
- package/ui/src/components/index.ts +0 -5
- package/ui/src/context/auth.tsx +0 -80
- package/ui/src/context/theme.tsx +0 -66
- package/ui/src/hooks/useAutoRefresh.ts +0 -39
- package/ui/src/i18n/ui.ts +0 -473
- package/ui/src/main.tsx +0 -16
- package/ui/src/pages/EvolutionPage.tsx +0 -333
- package/ui/src/pages/FeedbackPage.tsx +0 -138
- package/ui/src/pages/GateMonitorPage.tsx +0 -136
- package/ui/src/pages/LoginPage.tsx +0 -89
- package/ui/src/pages/OverviewPage.tsx +0 -599
- package/ui/src/pages/SamplesPage.tsx +0 -174
- package/ui/src/pages/ThinkingModelsPage.tsx +0 -702
- package/ui/src/styles.css +0 -2020
- package/ui/src/types.ts +0 -384
- package/ui/src/utils/format.ts +0 -15
|
@@ -1,234 +0,0 @@
|
|
|
1
|
-
import { describe, it, expect, beforeEach, afterEach } from 'vitest';
|
|
2
|
-
import * as fs from 'fs';
|
|
3
|
-
import * as os from 'os';
|
|
4
|
-
import * as path from 'path';
|
|
5
|
-
import { TrajectoryDatabase } from '../../src/core/trajectory.js';
|
|
6
|
-
import { NocturnalTrajectoryExtractor } from '../../src/core/nocturnal-trajectory-extractor.js';
|
|
7
|
-
import { detectViolation } from '../../src/core/nocturnal-compliance.js';
|
|
8
|
-
|
|
9
|
-
function safeRmDir(dir: string): void {
|
|
10
|
-
try { fs.rmSync(dir, { recursive: true, force: true }); } catch { /* ignore */ }
|
|
11
|
-
}
|
|
12
|
-
|
|
13
|
-
// ─────────────────────────────────────────────────────────
|
|
14
|
-
// Phase 4a: Correction rejected → pain event → nocturnal selection
|
|
15
|
-
// ─────────────────────────────────────────────────────────
|
|
16
|
-
describe('Phase 4a: Correction rejected integration', () => {
|
|
17
|
-
let workspaceDir: string;
|
|
18
|
-
let trajectory: TrajectoryDatabase;
|
|
19
|
-
|
|
20
|
-
beforeEach(() => {
|
|
21
|
-
workspaceDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pd-e2e-correction-'));
|
|
22
|
-
trajectory = new TrajectoryDatabase({ workspaceDir });
|
|
23
|
-
});
|
|
24
|
-
|
|
25
|
-
afterEach(() => {
|
|
26
|
-
trajectory?.dispose();
|
|
27
|
-
safeRmDir(workspaceDir);
|
|
28
|
-
});
|
|
29
|
-
|
|
30
|
-
it('rejected correction creates a pain event with source=correction_rejected', () => {
|
|
31
|
-
// 1. Create session + correction sample
|
|
32
|
-
trajectory.recordSession({ sessionId: 'corr-session', startedAt: new Date().toISOString() });
|
|
33
|
-
const atId = trajectory.recordAssistantTurn({
|
|
34
|
-
sessionId: 'corr-session', runId: 'run-1', provider: 'local', model: 'main',
|
|
35
|
-
rawText: 'Here is my code', sanitizedText: 'Here is my code', usageJson: {}, empathySignalJson: {},
|
|
36
|
-
createdAt: new Date().toISOString(),
|
|
37
|
-
});
|
|
38
|
-
trajectory.recordUserTurn({
|
|
39
|
-
sessionId: 'corr-session', turnIndex: 1, rawText: 'This is wrong!',
|
|
40
|
-
correctionDetected: true, correctionCue: '错了',
|
|
41
|
-
referencesAssistantTurnId: atId, createdAt: new Date().toISOString(),
|
|
42
|
-
});
|
|
43
|
-
// Tool call triggers maybeCreateCorrectionSample on success
|
|
44
|
-
trajectory.recordToolCall({
|
|
45
|
-
sessionId: 'corr-session', toolName: 'read', outcome: 'success',
|
|
46
|
-
createdAt: new Date().toISOString(),
|
|
47
|
-
});
|
|
48
|
-
|
|
49
|
-
// Verify sample was created
|
|
50
|
-
const samples = trajectory.listCorrectionSamples('pending');
|
|
51
|
-
expect(samples.length).toBe(1);
|
|
52
|
-
|
|
53
|
-
// 2. Reject the sample
|
|
54
|
-
trajectory.reviewCorrectionSample(samples[0].sampleId, 'rejected', 'Bad approach');
|
|
55
|
-
|
|
56
|
-
// 3. Verify pain event was created
|
|
57
|
-
const painEvents = trajectory.listPainEventsForSession('corr-session');
|
|
58
|
-
const correctionPain = painEvents.find(e => e.source === 'correction_rejected');
|
|
59
|
-
expect(correctionPain).toBeDefined();
|
|
60
|
-
expect(correctionPain!.score).toBeGreaterThanOrEqual(0);
|
|
61
|
-
expect(correctionPain!.score).toBeLessThanOrEqual(100);
|
|
62
|
-
});
|
|
63
|
-
|
|
64
|
-
it('approved correction does NOT create a pain event', () => {
|
|
65
|
-
trajectory.recordSession({ sessionId: 'approved-session', startedAt: new Date().toISOString() });
|
|
66
|
-
const atId = trajectory.recordAssistantTurn({
|
|
67
|
-
sessionId: 'approved-session', runId: 'run-2', provider: 'local', model: 'main',
|
|
68
|
-
rawText: 'Good code', sanitizedText: 'Good code', usageJson: {}, empathySignalJson: {},
|
|
69
|
-
createdAt: new Date().toISOString(),
|
|
70
|
-
});
|
|
71
|
-
trajectory.recordUserTurn({
|
|
72
|
-
sessionId: 'approved-session', turnIndex: 1, rawText: 'Looks better',
|
|
73
|
-
correctionDetected: true, correctionCue: '改进',
|
|
74
|
-
referencesAssistantTurnId: atId, createdAt: new Date().toISOString(),
|
|
75
|
-
});
|
|
76
|
-
// Tool call triggers maybeCreateCorrectionSample on success
|
|
77
|
-
trajectory.recordToolCall({
|
|
78
|
-
sessionId: 'approved-session', toolName: 'read', outcome: 'success',
|
|
79
|
-
createdAt: new Date().toISOString(),
|
|
80
|
-
});
|
|
81
|
-
|
|
82
|
-
const samples = trajectory.listCorrectionSamples('pending');
|
|
83
|
-
expect(samples.length).toBe(1);
|
|
84
|
-
|
|
85
|
-
trajectory.reviewCorrectionSample(samples[0].sampleId, 'approved', 'Good');
|
|
86
|
-
|
|
87
|
-
const painEvents = trajectory.listPainEventsForSession('approved-session');
|
|
88
|
-
const correctionPain = painEvents.find(e => e.source === 'correction_rejected');
|
|
89
|
-
expect(correctionPain).toBeUndefined();
|
|
90
|
-
});
|
|
91
|
-
});
|
|
92
|
-
|
|
93
|
-
// ─────────────────────────────────────────────────────────
|
|
94
|
-
// Phase 4b: Gate block + pain multi-signal test
|
|
95
|
-
// ─────────────────────────────────────────────────────────
|
|
96
|
-
describe('Phase 4b: Multi-signal session selection', () => {
|
|
97
|
-
let workspaceDir: string;
|
|
98
|
-
let trajectory: TrajectoryDatabase;
|
|
99
|
-
|
|
100
|
-
beforeEach(() => {
|
|
101
|
-
workspaceDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pd-e2e-multisignal-'));
|
|
102
|
-
trajectory = new TrajectoryDatabase({ workspaceDir });
|
|
103
|
-
});
|
|
104
|
-
|
|
105
|
-
afterEach(() => {
|
|
106
|
-
trajectory?.dispose();
|
|
107
|
-
safeRmDir(workspaceDir);
|
|
108
|
-
});
|
|
109
|
-
|
|
110
|
-
it('session with more failures has higher violation density', () => {
|
|
111
|
-
// Create session A: just 1 failure
|
|
112
|
-
trajectory.recordSession({ sessionId: 'session-a-pain-only', startedAt: new Date().toISOString() });
|
|
113
|
-
const atIdA = trajectory.recordAssistantTurn({
|
|
114
|
-
sessionId: 'session-a-pain-only', runId: 'run-a', provider: 'local', model: 'main',
|
|
115
|
-
rawText: 'Code here', sanitizedText: 'Code here', usageJson: {}, empathySignalJson: {},
|
|
116
|
-
createdAt: new Date().toISOString(),
|
|
117
|
-
});
|
|
118
|
-
trajectory.recordUserTurn({
|
|
119
|
-
sessionId: 'session-a-pain-only', turnIndex: 1, rawText: '错了',
|
|
120
|
-
correctionDetected: true, correctionCue: '错了',
|
|
121
|
-
referencesAssistantTurnId: atIdA, createdAt: new Date().toISOString(),
|
|
122
|
-
});
|
|
123
|
-
trajectory.recordToolCall({
|
|
124
|
-
sessionId: 'session-a-pain-only', toolName: 'write', outcome: 'failure',
|
|
125
|
-
errorMessage: 'Write failed', errorType: 'Error', createdAt: new Date().toISOString(),
|
|
126
|
-
});
|
|
127
|
-
|
|
128
|
-
// Create session B: 2 failures
|
|
129
|
-
trajectory.recordSession({ sessionId: 'session-b-multi', startedAt: new Date().toISOString() });
|
|
130
|
-
const atIdB = trajectory.recordAssistantTurn({
|
|
131
|
-
sessionId: 'session-b-multi', runId: 'run-b', provider: 'local', model: 'main',
|
|
132
|
-
rawText: 'Code here', sanitizedText: 'Code here', usageJson: {}, empathySignalJson: {},
|
|
133
|
-
createdAt: new Date().toISOString(),
|
|
134
|
-
});
|
|
135
|
-
trajectory.recordUserTurn({
|
|
136
|
-
sessionId: 'session-b-multi', turnIndex: 1, rawText: '太复杂了',
|
|
137
|
-
correctionDetected: true, correctionCue: '太复杂了',
|
|
138
|
-
referencesAssistantTurnId: atIdB, createdAt: new Date().toISOString(),
|
|
139
|
-
});
|
|
140
|
-
trajectory.recordToolCall({
|
|
141
|
-
sessionId: 'session-b-multi', toolName: 'edit', outcome: 'failure',
|
|
142
|
-
errorMessage: 'Edit failed', errorType: 'Error', createdAt: new Date().toISOString(),
|
|
143
|
-
});
|
|
144
|
-
trajectory.recordToolCall({
|
|
145
|
-
sessionId: 'session-b-multi', toolName: 'write', outcome: 'failure',
|
|
146
|
-
errorMessage: 'Write failed too', errorType: 'Error', createdAt: new Date().toISOString(),
|
|
147
|
-
});
|
|
148
|
-
|
|
149
|
-
// Verify session B has more failure signals
|
|
150
|
-
const extractor = new NocturnalTrajectoryExtractor(trajectory);
|
|
151
|
-
const snapshotA = extractor.getNocturnalSessionSnapshot('session-a-pain-only');
|
|
152
|
-
const snapshotB = extractor.getNocturnalSessionSnapshot('session-b-multi');
|
|
153
|
-
|
|
154
|
-
expect(snapshotA).not.toBeNull();
|
|
155
|
-
expect(snapshotB).not.toBeNull();
|
|
156
|
-
|
|
157
|
-
// Session B should have more violation signals
|
|
158
|
-
const densityA = (snapshotA!.stats.failureCount ?? 0) + (snapshotA!.stats.totalPainEvents ?? 0) * 0.5;
|
|
159
|
-
const densityB = (snapshotB!.stats.failureCount ?? 0) + (snapshotB!.stats.totalPainEvents ?? 0) * 0.5;
|
|
160
|
-
expect(densityB).toBeGreaterThan(densityA);
|
|
161
|
-
});
|
|
162
|
-
});
|
|
163
|
-
|
|
164
|
-
// ─────────────────────────────────────────────────────────
|
|
165
|
-
// Phase 4c: Boundary value test matrix
|
|
166
|
-
// ─────────────────────────────────────────────────────────
|
|
167
|
-
describe('Phase 4c: Boundary value tests', () => {
|
|
168
|
-
let workspaceDir: string;
|
|
169
|
-
let trajectory: TrajectoryDatabase;
|
|
170
|
-
|
|
171
|
-
beforeEach(() => {
|
|
172
|
-
workspaceDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pd-e2e-boundary-'));
|
|
173
|
-
trajectory = new TrajectoryDatabase({ workspaceDir });
|
|
174
|
-
});
|
|
175
|
-
|
|
176
|
-
afterEach(() => {
|
|
177
|
-
trajectory?.dispose();
|
|
178
|
-
safeRmDir(workspaceDir);
|
|
179
|
-
});
|
|
180
|
-
|
|
181
|
-
it('session with correction cue is listed as candidate', () => {
|
|
182
|
-
trajectory.recordSession({ sessionId: 'single-pain', startedAt: new Date().toISOString() });
|
|
183
|
-
const atIdC = trajectory.recordAssistantTurn({
|
|
184
|
-
sessionId: 'single-pain', runId: 'run-c', provider: 'local', model: 'main',
|
|
185
|
-
rawText: 'Agent response', sanitizedText: 'Agent response', usageJson: {}, empathySignalJson: {},
|
|
186
|
-
createdAt: new Date().toISOString(),
|
|
187
|
-
});
|
|
188
|
-
trajectory.recordUserTurn({
|
|
189
|
-
sessionId: 'single-pain', turnIndex: 1, rawText: '错了',
|
|
190
|
-
correctionDetected: true, correctionCue: '错了',
|
|
191
|
-
referencesAssistantTurnId: atIdC, createdAt: new Date().toISOString(),
|
|
192
|
-
});
|
|
193
|
-
|
|
194
|
-
const extractor = new NocturnalTrajectoryExtractor(trajectory);
|
|
195
|
-
const candidates = extractor.listRecentNocturnalCandidateSessions({ limit: 10, minToolCalls: 0 });
|
|
196
|
-
|
|
197
|
-
const painCandidate = candidates.find(c => c.sessionId === 'single-pain');
|
|
198
|
-
expect(painCandidate).toBeDefined();
|
|
199
|
-
});
|
|
200
|
-
|
|
201
|
-
it('detectViolation returns violated for P_* principles with tool failure', () => {
|
|
202
|
-
trajectory.recordSession({ sessionId: 'violation-session', startedAt: new Date().toISOString() });
|
|
203
|
-
trajectory.recordAssistantTurn({
|
|
204
|
-
sessionId: 'violation-session', runId: 'run-d', provider: 'local', model: 'main',
|
|
205
|
-
rawText: 'Code', sanitizedText: 'Code', usageJson: {}, empathySignalJson: {},
|
|
206
|
-
createdAt: new Date().toISOString(),
|
|
207
|
-
});
|
|
208
|
-
trajectory.recordToolCall({
|
|
209
|
-
sessionId: 'violation-session', toolName: 'write', outcome: 'failure',
|
|
210
|
-
errorMessage: 'Failed', errorType: 'Error', createdAt: new Date().toISOString(),
|
|
211
|
-
});
|
|
212
|
-
|
|
213
|
-
const extractor = new NocturnalTrajectoryExtractor(trajectory);
|
|
214
|
-
const snapshot = extractor.getNocturnalSessionSnapshot('violation-session');
|
|
215
|
-
expect(snapshot).not.toBeNull();
|
|
216
|
-
|
|
217
|
-
// P_* principles should be violated with any failure
|
|
218
|
-
const violation = detectViolation('P_001', {
|
|
219
|
-
sessionId: 'violation-session',
|
|
220
|
-
toolCalls: snapshot!.toolCalls.map(tc => ({
|
|
221
|
-
toolName: tc.toolName, outcome: tc.outcome as 'success' | 'failure' | 'blocked',
|
|
222
|
-
errorMessage: tc.errorMessage ?? undefined,
|
|
223
|
-
})),
|
|
224
|
-
painSignals: snapshot!.painEvents.map(pe => ({
|
|
225
|
-
source: pe.source, score: pe.score, severity: pe.severity as 'mild' | 'moderate' | 'severe' | undefined,
|
|
226
|
-
})),
|
|
227
|
-
gateBlocks: [],
|
|
228
|
-
userCorrections: [],
|
|
229
|
-
planApprovals: [],
|
|
230
|
-
});
|
|
231
|
-
|
|
232
|
-
expect(violation.violated).toBe(true);
|
|
233
|
-
});
|
|
234
|
-
});
|
|
@@ -1,357 +0,0 @@
|
|
|
1
|
-
import { describe, it, expect } from 'vitest';
|
|
2
|
-
import {
|
|
3
|
-
validateExecutability,
|
|
4
|
-
validateForApproval,
|
|
5
|
-
type BoundedAction,
|
|
6
|
-
} from '../../src/core/nocturnal-executability.js';
|
|
7
|
-
|
|
8
|
-
describe('Nocturnal Executability', () => {
|
|
9
|
-
|
|
10
|
-
// -------------------------------------------------------------------------
|
|
11
|
-
// Tests: validateExecutability — valid bounded actions
|
|
12
|
-
// -------------------------------------------------------------------------
|
|
13
|
-
|
|
14
|
-
describe('validateExecutability — valid bounded actions', () => {
|
|
15
|
-
it('accepts "Read the file before editing"', () => {
|
|
16
|
-
const result = validateExecutability({
|
|
17
|
-
badDecision: 'Edited a file without reading it first',
|
|
18
|
-
betterDecision: 'Read the file before editing to understand its current structure',
|
|
19
|
-
});
|
|
20
|
-
expect(result.executable).toBe(true);
|
|
21
|
-
expect(result.boundedAction).toBeDefined();
|
|
22
|
-
expect(result.boundedAction?.verb.toLowerCase()).toBe('read');
|
|
23
|
-
});
|
|
24
|
-
|
|
25
|
-
it('accepts "Check the error message and verify preconditions"', () => {
|
|
26
|
-
const result = validateExecutability({
|
|
27
|
-
badDecision: 'Retried without checking error',
|
|
28
|
-
betterDecision: 'Check the error message and verify preconditions before retrying',
|
|
29
|
-
});
|
|
30
|
-
expect(result.executable).toBe(true);
|
|
31
|
-
expect(result.boundedAction?.verb.toLowerCase()).toBe('check');
|
|
32
|
-
});
|
|
33
|
-
|
|
34
|
-
it('accepts "Verify the current state of the repository"', () => {
|
|
35
|
-
const result = validateExecutability({
|
|
36
|
-
badDecision: 'Committed without checking status',
|
|
37
|
-
betterDecision: 'Verify the current state of the repository before committing',
|
|
38
|
-
});
|
|
39
|
-
expect(result.executable).toBe(true);
|
|
40
|
-
expect(result.boundedAction?.verb.toLowerCase()).toBe('verify');
|
|
41
|
-
});
|
|
42
|
-
|
|
43
|
-
it('accepts "Edit the config file to add the missing setting"', () => {
|
|
44
|
-
const result = validateExecutability({
|
|
45
|
-
badDecision: 'Changed settings without knowing current values',
|
|
46
|
-
betterDecision: 'Edit the config file to add the missing setting',
|
|
47
|
-
});
|
|
48
|
-
expect(result.executable).toBe(true);
|
|
49
|
-
expect(result.boundedAction?.verb.toLowerCase()).toBe('edit');
|
|
50
|
-
});
|
|
51
|
-
|
|
52
|
-
it('accepts "Search the codebase for similar patterns"', () => {
|
|
53
|
-
const result = validateExecutability({
|
|
54
|
-
badDecision: 'Wrote duplicate code',
|
|
55
|
-
betterDecision: 'Search the codebase for similar patterns before implementing new code',
|
|
56
|
-
});
|
|
57
|
-
expect(result.executable).toBe(true);
|
|
58
|
-
expect(result.boundedAction?.verb.toLowerCase()).toBe('search');
|
|
59
|
-
});
|
|
60
|
-
|
|
61
|
-
it('accepts "Look at the file to understand its structure"', () => {
|
|
62
|
-
const result = validateExecutability({
|
|
63
|
-
badDecision: 'Made changes without understanding the code',
|
|
64
|
-
betterDecision: 'Look at the file to understand its structure before modifying it',
|
|
65
|
-
});
|
|
66
|
-
expect(result.executable).toBe(true);
|
|
67
|
-
expect(result.boundedAction?.verb.toLowerCase()).toBe('look');
|
|
68
|
-
});
|
|
69
|
-
});
|
|
70
|
-
|
|
71
|
-
// -------------------------------------------------------------------------
|
|
72
|
-
// Tests: validateExecutability — vague verbs
|
|
73
|
-
// -------------------------------------------------------------------------
|
|
74
|
-
|
|
75
|
-
describe('validateExecutability — vague verbs', () => {
|
|
76
|
-
const vagueVerbTests = [
|
|
77
|
-
{ text: 'Understand the error before proceeding', expectedFail: true },
|
|
78
|
-
{ text: 'Learn from the failure and adjust', expectedFail: true },
|
|
79
|
-
{ text: 'Improve the error handling approach', expectedFail: true },
|
|
80
|
-
{ text: 'Fix the issue by checking preconditions', expectedFail: true }, // "fix" is vague
|
|
81
|
-
{ text: 'Handle errors more gracefully', expectedFail: true },
|
|
82
|
-
{ text: 'Be more careful with edge cases', expectedFail: true },
|
|
83
|
-
{ text: 'Ensure the code is correct', expectedFail: true },
|
|
84
|
-
{ text: 'Consider reading the documentation', expectedFail: true },
|
|
85
|
-
{ text: 'Think about the root cause', expectedFail: true },
|
|
86
|
-
{ text: 'Reflect on the failure', expectedFail: true },
|
|
87
|
-
{ text: 'Review the error carefully', expectedFail: true }, // borderline but rejected by our list
|
|
88
|
-
];
|
|
89
|
-
|
|
90
|
-
vagueVerbTests.forEach(({ text, expectedFail }) => {
|
|
91
|
-
it(`"${text}" → ${expectedFail ? 'rejected' : 'accepted'}`, () => {
|
|
92
|
-
const result = validateExecutability({
|
|
93
|
-
badDecision: 'Made a mistake',
|
|
94
|
-
betterDecision: text,
|
|
95
|
-
});
|
|
96
|
-
if (expectedFail) {
|
|
97
|
-
expect(result.executable).toBe(false);
|
|
98
|
-
expect(result.failures.some(f => f.reason.includes('vague verb'))).toBe(true);
|
|
99
|
-
} else {
|
|
100
|
-
expect(result.executable).toBe(true);
|
|
101
|
-
}
|
|
102
|
-
});
|
|
103
|
-
});
|
|
104
|
-
});
|
|
105
|
-
|
|
106
|
-
// -------------------------------------------------------------------------
|
|
107
|
-
// Tests: validateExecutability — hollow patterns
|
|
108
|
-
// -------------------------------------------------------------------------
|
|
109
|
-
|
|
110
|
-
describe('validateExecutability — hollow patterns', () => {
|
|
111
|
-
const hollowTests = [
|
|
112
|
-
{ text: 'Always be careful when editing files', expectedFail: true },
|
|
113
|
-
{ text: 'Be mindful of potential conflicts', expectedFail: true },
|
|
114
|
-
{ text: "Don't rush into making changes", expectedFail: true },
|
|
115
|
-
{ text: 'Take your time with complex tasks', expectedFail: true },
|
|
116
|
-
{ text: 'Be patient and verify your work', expectedFail: true },
|
|
117
|
-
{ text: 'Be more careful next time', expectedFail: true },
|
|
118
|
-
{ text: 'Work smarter, not harder', expectedFail: true },
|
|
119
|
-
{ text: 'Follow best practices for error handling', expectedFail: true },
|
|
120
|
-
{ text: 'Read the documentation before proceeding', expectedFail: false }, // not hollow
|
|
121
|
-
];
|
|
122
|
-
|
|
123
|
-
hollowTests.forEach(({ text, expectedFail }) => {
|
|
124
|
-
it(`"${text}" → ${expectedFail ? 'rejected (hollow)' : 'accepted'}`, () => {
|
|
125
|
-
const result = validateExecutability({
|
|
126
|
-
badDecision: 'Made an error',
|
|
127
|
-
betterDecision: text,
|
|
128
|
-
});
|
|
129
|
-
if (expectedFail) {
|
|
130
|
-
expect(result.executable).toBe(false);
|
|
131
|
-
expect(result.failures.some(f => f.reason.includes('hollow'))).toBe(true);
|
|
132
|
-
} else {
|
|
133
|
-
expect(result.executable).toBe(true);
|
|
134
|
-
}
|
|
135
|
-
});
|
|
136
|
-
});
|
|
137
|
-
});
|
|
138
|
-
|
|
139
|
-
// -------------------------------------------------------------------------
|
|
140
|
-
// Tests: validateExecutability — too generic
|
|
141
|
-
// -------------------------------------------------------------------------
|
|
142
|
-
|
|
143
|
-
describe('validateExecutability — too generic', () => {
|
|
144
|
-
it('rejects text that is too short', () => {
|
|
145
|
-
const result = validateExecutability({
|
|
146
|
-
badDecision: 'Made a mistake',
|
|
147
|
-
betterDecision: 'Be better',
|
|
148
|
-
});
|
|
149
|
-
expect(result.executable).toBe(false);
|
|
150
|
-
expect(result.failures.some(f => f.reason.includes('too generic'))).toBe(true);
|
|
151
|
-
});
|
|
152
|
-
|
|
153
|
-
it('rejects purely negative constraints', () => {
|
|
154
|
-
const result = validateExecutability({
|
|
155
|
-
badDecision: 'Did not check the error',
|
|
156
|
-
betterDecision: 'Do not make the same mistake again',
|
|
157
|
-
});
|
|
158
|
-
expect(result.executable).toBe(false);
|
|
159
|
-
expect(result.failures.some(f => f.reason.includes('too generic'))).toBe(true);
|
|
160
|
-
});
|
|
161
|
-
|
|
162
|
-
it('accepts a concrete action with target', () => {
|
|
163
|
-
const result = validateExecutability({
|
|
164
|
-
badDecision: 'Edited without reading',
|
|
165
|
-
betterDecision: 'Read src/main.ts to understand the current structure before editing',
|
|
166
|
-
});
|
|
167
|
-
expect(result.executable).toBe(true);
|
|
168
|
-
});
|
|
169
|
-
});
|
|
170
|
-
|
|
171
|
-
// -------------------------------------------------------------------------
|
|
172
|
-
// Tests: validateExecutability — non-existent operations
|
|
173
|
-
// -------------------------------------------------------------------------
|
|
174
|
-
|
|
175
|
-
describe('validateExecutability — non-existent operations', () => {
|
|
176
|
-
it('rejects "rewrite the entire codebase"', () => {
|
|
177
|
-
const result = validateExecutability({
|
|
178
|
-
badDecision: 'Made changes poorly',
|
|
179
|
-
betterDecision: 'Rewrite the entire codebase to fix the architecture',
|
|
180
|
-
});
|
|
181
|
-
expect(result.executable).toBe(false);
|
|
182
|
-
expect(result.failures.some(f => f.reason.includes('too broad'))).toBe(true);
|
|
183
|
-
});
|
|
184
|
-
|
|
185
|
-
it('rejects "redesign the whole system"', () => {
|
|
186
|
-
const result = validateExecutability({
|
|
187
|
-
badDecision: 'System design was flawed',
|
|
188
|
-
betterDecision: 'Redesign the whole system from scratch',
|
|
189
|
-
});
|
|
190
|
-
expect(result.executable).toBe(false);
|
|
191
|
-
expect(result.failures.some(f => f.reason.includes('too broad'))).toBe(true);
|
|
192
|
-
});
|
|
193
|
-
|
|
194
|
-
it('rejects "restart from scratch"', () => {
|
|
195
|
-
const result = validateExecutability({
|
|
196
|
-
badDecision: 'Initial approach was wrong',
|
|
197
|
-
betterDecision: 'Restart from scratch with a new plan',
|
|
198
|
-
});
|
|
199
|
-
expect(result.executable).toBe(false);
|
|
200
|
-
expect(result.failures.some(f => f.reason.includes('too broad'))).toBe(true);
|
|
201
|
-
});
|
|
202
|
-
});
|
|
203
|
-
|
|
204
|
-
// -------------------------------------------------------------------------
|
|
205
|
-
// Tests: validateExecutability — badDecision hollow check
|
|
206
|
-
// -------------------------------------------------------------------------
|
|
207
|
-
|
|
208
|
-
describe('validateExecutability — badDecision checks', () => {
|
|
209
|
-
it('rejects hollow badDecision pattern', () => {
|
|
210
|
-
const result = validateExecutability({
|
|
211
|
-
badDecision: 'Always be careful and mindful of your actions',
|
|
212
|
-
betterDecision: 'Read the file before editing',
|
|
213
|
-
});
|
|
214
|
-
// The hollow check applies to badDecision too
|
|
215
|
-
expect(result.executable).toBe(false);
|
|
216
|
-
expect(result.failures.some(f => f.field === 'badDecision')).toBe(true);
|
|
217
|
-
});
|
|
218
|
-
|
|
219
|
-
it('accepts specific badDecision', () => {
|
|
220
|
-
const result = validateExecutability({
|
|
221
|
-
badDecision: 'Edited a file without reading it first, causing a merge conflict',
|
|
222
|
-
betterDecision: 'Read the file before editing to check for conflicts',
|
|
223
|
-
});
|
|
224
|
-
expect(result.executable).toBe(true);
|
|
225
|
-
});
|
|
226
|
-
});
|
|
227
|
-
|
|
228
|
-
// -------------------------------------------------------------------------
|
|
229
|
-
// Tests: validateForApproval — combined arbiter + executability
|
|
230
|
-
// -------------------------------------------------------------------------
|
|
231
|
-
|
|
232
|
-
describe('validateForApproval — combined check', () => {
|
|
233
|
-
function makeValidArtifact(overrides: Record<string, unknown> = {}): string {
|
|
234
|
-
return JSON.stringify({
|
|
235
|
-
artifactId: 'a1b2c3d4-e5f6-7890-abcd-ef1234567890',
|
|
236
|
-
sessionId: 'session-abc123',
|
|
237
|
-
principleId: 'T-08',
|
|
238
|
-
sourceSnapshotRef: 'snapshot-001',
|
|
239
|
-
badDecision: 'After bash command failed, immediately retried without diagnosing the root cause',
|
|
240
|
-
betterDecision: 'Check the error message and verify preconditions before retrying a failed bash command',
|
|
241
|
-
rationale: 'Treating each failure as a signal to diagnose rather than blindly retry prevents repeated failures and respects the cost of each action',
|
|
242
|
-
createdAt: '2026-03-27T12:00:00.000Z',
|
|
243
|
-
...overrides,
|
|
244
|
-
});
|
|
245
|
-
}
|
|
246
|
-
|
|
247
|
-
it('approves a valid artifact that passes both arbiter and executability', () => {
|
|
248
|
-
const json = makeValidArtifact();
|
|
249
|
-
const result = validateForApproval(json);
|
|
250
|
-
expect(result.approved).toBe(true);
|
|
251
|
-
expect(result.artifact).toBeDefined();
|
|
252
|
-
expect(result.failures).toHaveLength(0);
|
|
253
|
-
});
|
|
254
|
-
|
|
255
|
-
it('rejects artifact with invalid JSON', () => {
|
|
256
|
-
const result = validateForApproval('not json');
|
|
257
|
-
expect(result.approved).toBe(false);
|
|
258
|
-
expect(result.failures.some(f => f.includes('parse'))).toBe(true);
|
|
259
|
-
});
|
|
260
|
-
|
|
261
|
-
it('rejects artifact with missing principleId', () => {
|
|
262
|
-
const json = makeValidArtifact({ principleId: undefined });
|
|
263
|
-
const result = validateForApproval(json);
|
|
264
|
-
expect(result.approved).toBe(false);
|
|
265
|
-
});
|
|
266
|
-
|
|
267
|
-
it('rejects artifact with cross-validation mismatch', () => {
|
|
268
|
-
const json = makeValidArtifact({ principleId: 'T-08' });
|
|
269
|
-
const result = validateForApproval(json, { expectedPrincipleId: 'T-01' });
|
|
270
|
-
expect(result.approved).toBe(false);
|
|
271
|
-
expect(result.failures.some(f => f.includes('mismatch'))).toBe(true);
|
|
272
|
-
});
|
|
273
|
-
|
|
274
|
-
it('rejects artifact with vague verb in betterDecision', () => {
|
|
275
|
-
const json = makeValidArtifact({ betterDecision: 'Understand the error first' });
|
|
276
|
-
const result = validateForApproval(json);
|
|
277
|
-
expect(result.approved).toBe(false);
|
|
278
|
-
expect(result.failures.some(f => f.includes('vague verb'))).toBe(true);
|
|
279
|
-
});
|
|
280
|
-
|
|
281
|
-
it('rejects artifact with hollow pattern in betterDecision', () => {
|
|
282
|
-
const json = makeValidArtifact({ betterDecision: 'Always be careful when proceeding' });
|
|
283
|
-
const result = validateForApproval(json);
|
|
284
|
-
expect(result.approved).toBe(false);
|
|
285
|
-
expect(result.failures.some(f => f.includes('hollow'))).toBe(true);
|
|
286
|
-
});
|
|
287
|
-
|
|
288
|
-
it('rejects artifact with invalid: true', () => {
|
|
289
|
-
const json = makeValidArtifact({ invalid: true, reason: 'no violation found' });
|
|
290
|
-
const result = validateForApproval(json);
|
|
291
|
-
expect(result.approved).toBe(false);
|
|
292
|
-
});
|
|
293
|
-
|
|
294
|
-
it('returns boundedAction in approved artifact', () => {
|
|
295
|
-
const json = makeValidArtifact();
|
|
296
|
-
const result = validateForApproval(json);
|
|
297
|
-
expect(result.approved).toBe(true);
|
|
298
|
-
expect(result.artifact?.boundedAction).toBeDefined();
|
|
299
|
-
expect(result.artifact?.boundedAction?.verb.toLowerCase()).toBe('check');
|
|
300
|
-
});
|
|
301
|
-
|
|
302
|
-
it('passes when cross-validation matches', () => {
|
|
303
|
-
const json = makeValidArtifact({ principleId: 'T-08', sessionId: 'session-xyz' });
|
|
304
|
-
const result = validateForApproval(json, {
|
|
305
|
-
expectedPrincipleId: 'T-08',
|
|
306
|
-
expectedSessionId: 'session-xyz',
|
|
307
|
-
});
|
|
308
|
-
expect(result.approved).toBe(true);
|
|
309
|
-
});
|
|
310
|
-
});
|
|
311
|
-
|
|
312
|
-
// -------------------------------------------------------------------------
|
|
313
|
-
// Tests: bounded action parsing
|
|
314
|
-
// -------------------------------------------------------------------------
|
|
315
|
-
|
|
316
|
-
describe('bounded action parsing', () => {
|
|
317
|
-
it('extracts verb and target from "Read the file"', () => {
|
|
318
|
-
const result = validateExecutability({
|
|
319
|
-
badDecision: 'Did not read',
|
|
320
|
-
betterDecision: 'Read the file before proceeding',
|
|
321
|
-
});
|
|
322
|
-
expect(result.executable).toBe(true);
|
|
323
|
-
expect(result.boundedAction?.verb.toLowerCase()).toBe('read');
|
|
324
|
-
// The boundedPattern captures (?:word\s+){0,4} after first word, so target = "file before proceeding"
|
|
325
|
-
expect(result.boundedAction?.target).toBe('file before proceeding');
|
|
326
|
-
});
|
|
327
|
-
|
|
328
|
-
it('extracts verb and target from "Check X first"', () => {
|
|
329
|
-
const result = validateExecutability({
|
|
330
|
-
badDecision: 'Did not check',
|
|
331
|
-
betterDecision: 'Check the error message first',
|
|
332
|
-
});
|
|
333
|
-
expect(result.executable).toBe(true);
|
|
334
|
-
expect(result.boundedAction?.verb.toLowerCase()).toBe('check');
|
|
335
|
-
// "Check the error message first" → verb="check", target="error message first"
|
|
336
|
-
expect(result.boundedAction?.target).toBe('error message first');
|
|
337
|
-
});
|
|
338
|
-
|
|
339
|
-
it('includes fullText in bounded action', () => {
|
|
340
|
-
const result = validateExecutability({
|
|
341
|
-
badDecision: 'Made a mistake',
|
|
342
|
-
betterDecision: 'Verify the preconditions before running the command',
|
|
343
|
-
});
|
|
344
|
-
expect(result.executable).toBe(true);
|
|
345
|
-
expect(result.boundedAction?.fullText.toLowerCase()).toContain('verify');
|
|
346
|
-
});
|
|
347
|
-
|
|
348
|
-
it('returns executable: false without boundedAction when all checks fail', () => {
|
|
349
|
-
const result = validateExecutability({
|
|
350
|
-
badDecision: 'Made a mistake',
|
|
351
|
-
betterDecision: 'Be better next time',
|
|
352
|
-
});
|
|
353
|
-
expect(result.executable).toBe(false);
|
|
354
|
-
expect(result.boundedAction).toBeUndefined();
|
|
355
|
-
});
|
|
356
|
-
});
|
|
357
|
-
});
|