principles-disciple 1.8.1 → 1.8.3

package/dist/core/evolution-types.d.ts

@@ -1,287 +0,0 @@
-/**
- * Evolution Points System V2.0 - MVP
- *
- * Core Philosophy: Growth-driven替代Penalty-driven
- * - 起点0分，只能增加，不扣分
- * - 失败记录教训，不扣分
- * - 同类任务失败后首次成功 = 双倍奖励（1小时冷却）
- * - 5级成长路径：Seed → Forest
- */
-export declare enum EvolutionTier {
-    Seed = 1,// 起步：150行 + 3文件 + 子智能体（现代 AI 能力已足够强）
-    Sprout = 2,// 成长：300行 + 5文件
-    Sapling = 3,// 独当：500行 + 10文件 + 风险路径
-    Tree = 4,// 专家：1000行 + 20文件
-    Forest = 5
-}
-export interface TierPermissions {
-    maxLinesPerWrite: number;
-    maxFilesPerTask: number;
-    allowRiskPath: boolean;
-    allowSubagentSpawn: boolean;
-}
-export interface TierDefinition {
-    tier: EvolutionTier;
-    name: string;
-    requiredPoints: number;
-    permissions: TierPermissions;
-}
-export declare const TIER_DEFINITIONS: readonly TierDefinition[];
-export declare function getTierDefinition(tier: EvolutionTier): TierDefinition;
-export declare function getTierByPoints(totalPoints: number): EvolutionTier;
-export type TaskDifficulty = 'trivial' | 'normal' | 'hard';
-export interface TaskDifficultyConfig {
-    basePoints: number;
-    description: string;
-}
-export declare const TASK_DIFFICULTY_CONFIG: Record<TaskDifficulty, TaskDifficultyConfig>;
-export type EvolutionEventType = 'success' | 'failure';
-export interface EvolutionEvent {
-    id: string;
-    timestamp: string;
-    type: EvolutionEventType;
-    taskHash: string;
-    taskDifficulty: TaskDifficulty;
-    toolName?: string;
-    filePath?: string;
-    reason?: string;
-    pointsAwarded: number;
-    isDoubleReward: boolean;
-    sessionId?: string;
-}
-export interface EvolutionScorecard {
-    version: '2.0';
-    agentId: string;
-    totalPoints: number;
-    availablePoints: number;
-    currentTier: EvolutionTier;
-    lastDoubleRewardTime?: string;
-    recentFailureHashes: Map<string, string>;
-    stats: EvolutionStats;
-    recentEvents: EvolutionEvent[];
-    lastUpdated: string;
-}
-export interface EvolutionStats {
-    totalSuccesses: number;
-    totalFailures: number;
-    consecutiveSuccesses: number;
-    consecutiveFailures: number;
-    doubleRewardsEarned: number;
-    tierPromotions: number;
-    pointsByDifficulty: Record<TaskDifficulty, number>;
-}
-export interface EvolutionStorage {
-    scorecard: EvolutionScorecard;
-    archivedStats: {
-        totalEventsProcessed: number;
-        pointsFromTrivial: number;
-        pointsFromNormal: number;
-        pointsFromHard: number;
-    };
-}
-export interface EvolutionConfig {
-    /** 双倍奖励冷却时间（毫秒），默认1小时 */
-    doubleRewardCooldownMs: number;
-    /** 保存的最近事件数量，默认50 */
-    maxRecentEvents: number;
-    /** 高等级做低级任务的积分衰减系数 */
-    difficultyPenalty: {
-        tier4Trivial: number;
-        tier4Normal: number;
-        tier5Trivial: number;
-        tier5Normal: number;
-    };
-}
-export declare const DEFAULT_EVOLUTION_CONFIG: EvolutionConfig;
-export interface ArchivedEventStats {
-    totalEventsProcessed: number;
-    pointsFromTrivial: number;
-    pointsFromNormal: number;
-    pointsFromHard: number;
-}
-export interface GateDecision {
-    allowed: boolean;
-    reason?: string;
-    currentTier?: EvolutionTier;
-    requiredTier?: EvolutionTier;
-}
-export interface ToolCallContext {
-    toolName: string;
-    filePath?: string;
-    content?: string;
-    lineCount?: number;
-    isRiskPath?: boolean;
-}
-export interface TierPromotionEvent {
-    previousTier: EvolutionTier;
-    newTier: EvolutionTier;
-    totalPoints: number;
-    timestamp: string;
-    newPermissions: TierPermissions;
-}
-export type PrincipleStatus = 'candidate' | 'probation' | 'active' | 'deprecated';
-/**
- * Evaluability classification — determines whether a P_xxx principle can enter
- * automatic nocturnal targeting.
- *
- * - deterministic:    Machine-checkable via deterministic rules/tool detectors
- * - weak_heuristic:  Checkable via heuristic signals, may have false positives
- * - manual_only:     No machine-checkable detector — stays in prompts only
- */
-export type PrincipleEvaluatorLevel = 'deterministic' | 'weak_heuristic' | 'manual_only';
-/**
- * Shared alias for PrincipleEvaluatorLevel — used by modules that reference
- * the same evaluability classification without direct coupling to evolution-types.
- * @deprecated Use PrincipleEvaluatorLevel directly. This alias exists for
- * backwards compatibility with principle-training-state.ts.
- */
-export type Evaluability = PrincipleEvaluatorLevel;
-/**
- * Structured detector metadata for P_xxx principles.
- * Allows the principle to enter automatic nocturnal targeting.
- *
- * If any required field is missing, the principle defaults to 'manual_only'.
- */
-export interface PrincipleDetectorSpec {
-    /** Topic/scenario tags where this detector applies */
-    applicabilityTags: string[];
-    /** Evidence that the principle was followed */
-    positiveSignals: string[];
-    /** Evidence that the principle was violated */
-    negativeSignals: string[];
-    /** Tool call sequences that indicate the principle is relevant */
-    toolSequenceHints: string[][];
-    /** Confidence in the detector's signal quality */
-    confidence: 'high' | 'medium' | 'low';
-}
-/**
- * Validates that a detector metadata object has all required fields with non-empty values.
- * Used as defense-in-depth before accepting detectorMetadata for auto-trainable principles.
- */
-export declare function isCompleteDetectorMetadata(meta: unknown): meta is PrincipleDetectorSpec;
-export interface Principle {
-    id: string;
-    version: number;
-    text: string;
-    source: {
-        painId: string;
-        painType: 'tool_failure' | 'subagent_error' | 'user_frustration';
-        timestamp: string;
-    };
-    trigger: string;
-    action: string;
-    guardrails?: string[];
-    contextTags: string[];
-    validation: {
-        successCount: number;
-        conflictCount: number;
-    };
-    status: PrincipleStatus;
-    feedbackScore: number;
-    usageCount: number;
-    createdAt: string;
-    activatedAt?: string;
-    deprecatedAt?: string;
-    /**
-     * Evaluability classification. Defaults to 'manual_only' if not set.
-     * Principles with 'manual_only' evaluability cannot enter automatic
-     * nocturnal targeting.
-     */
-    evaluability: PrincipleEvaluatorLevel;
-    /**
-     * Structured detector metadata. If present and valid, the principle
-     * may be auto-trainable (deterministic / weak_heuristic).
-     * Absent or malformed = 'manual_only' evaluability.
-     */
-    detectorMetadata?: PrincipleDetectorSpec;
-}
-export type EvolutionLoopEventType = 'pain_detected' | 'candidate_created' | 'principle_promoted' | 'principle_deprecated' | 'principle_rolled_back' | 'circuit_breaker_opened' | 'legacy_import';
-export interface PainDetectedData {
-    painId: string;
-    painType: 'tool_failure' | 'subagent_error' | 'user_frustration';
-    source: string;
-    reason: string;
-    score?: number;
-    sessionId?: string;
-    agentId?: string;
-    taskId?: string;
-    traceId?: string;
-}
-export interface CandidateCreatedData {
-    painId: string;
-    principleId: string;
-    trigger: string;
-    action: string;
-    status: 'candidate';
-    /** Pain type that generated this candidate — preserved on replay */
-    painType?: 'tool_failure' | 'subagent_error' | 'user_frustration';
-    /** Optional evaluability — defaults to 'manual_only' if omitted */
-    evaluability?: PrincipleEvaluatorLevel;
-    /** Optional detector metadata — absent = manual_only */
-    detectorMetadata?: PrincipleDetectorSpec;
-}
-export interface PrinciplePromotedData {
-    principleId: string;
-    from: 'candidate' | 'probation';
-    to: 'probation' | 'active';
-    reason: string;
-    successCount?: number;
-}
-export interface PrincipleDeprecatedData {
-    principleId: string;
-    reason: string;
-    triggeredBy: 'auto' | 'manual';
-}
-export interface PrincipleRolledBackData {
-    principleId: string;
-    reason: string;
-    triggeredBy: 'user_command' | 'auto_conflict';
-    blacklistPattern?: string;
-    relatedPainId?: string;
-}
-export interface CircuitBreakerOpenedData {
-    taskId: string;
-    painId: string;
-    failCount: number;
-    reason: string;
-    requireHuman: boolean;
-    nextRetryAt?: string;
-}
-export interface LegacyImportData {
-    sourceFile: string;
-    content: string;
-    contentHash?: string;
-}
-export type EvolutionLoopEvent = {
-    ts: string;
-    type: 'pain_detected';
-    data: PainDetectedData;
-} | {
-    ts: string;
-    type: 'pain_recorded';
-    data: PainDetectedData;
-} | {
-    ts: string;
-    type: 'candidate_created';
-    data: CandidateCreatedData;
-} | {
-    ts: string;
-    type: 'principle_promoted';
-    data: PrinciplePromotedData;
-} | {
-    ts: string;
-    type: 'principle_deprecated';
-    data: PrincipleDeprecatedData;
-} | {
-    ts: string;
-    type: 'principle_rolled_back';
-    data: PrincipleRolledBackData;
-} | {
-    ts: string;
-    type: 'circuit_breaker_opened';
-    data: CircuitBreakerOpenedData;
-} | {
-    ts: string;
-    type: 'legacy_import';
-    data: LegacyImportData;
-};

package/dist/core/evolution-types.js

@@ -1,78 +0,0 @@
-/**
- * Evolution Points System V2.0 - MVP
- *
- * Core Philosophy: Growth-driven替代Penalty-driven
- * - 起点0分，只能增加，不扣分
- * - 失败记录教训，不扣分
- * - 同类任务失败后首次成功 = 双倍奖励（1小时冷却）
- * - 5级成长路径：Seed → Forest
- */
-// ===== 等级定义 =====
-export var EvolutionTier;
-(function (EvolutionTier) {
-    EvolutionTier[EvolutionTier["Seed"] = 1] = "Seed";
-    EvolutionTier[EvolutionTier["Sprout"] = 2] = "Sprout";
-    EvolutionTier[EvolutionTier["Sapling"] = 3] = "Sapling";
-    EvolutionTier[EvolutionTier["Tree"] = 4] = "Tree";
-    EvolutionTier[EvolutionTier["Forest"] = 5] = "Forest"; // 大师：完全自主
-})(EvolutionTier || (EvolutionTier = {}));
-export const TIER_DEFINITIONS = [
-    // 2026-03-28: 大幅放宽限制，现代 AI 能力已很强
-    // Seed: 从 20 行提升到 150 行，允许 3 文件，允许子智能体
-    { tier: EvolutionTier.Seed, name: 'Seed', requiredPoints: 0, permissions: { maxLinesPerWrite: 150, maxFilesPerTask: 3, allowRiskPath: false, allowSubagentSpawn: true } },
-    // Sprout: 中等规模开发
-    { tier: EvolutionTier.Sprout, name: 'Sprout', requiredPoints: 50, permissions: { maxLinesPerWrite: 300, maxFilesPerTask: 5, allowRiskPath: false, allowSubagentSpawn: true } },
-    // Sapling: 较大规模开发，可访问风险路径（需 PLAN）
-    { tier: EvolutionTier.Sapling, name: 'Sapling', requiredPoints: 200, permissions: { maxLinesPerWrite: 500, maxFilesPerTask: 10, allowRiskPath: true, allowSubagentSpawn: true } },
-    // Tree: 大型重构
-    { tier: EvolutionTier.Tree, name: 'Tree', requiredPoints: 500, permissions: { maxLinesPerWrite: 1000, maxFilesPerTask: 20, allowRiskPath: true, allowSubagentSpawn: true } },
-    // Forest: 完全自主
-    { tier: EvolutionTier.Forest, name: 'Forest', requiredPoints: 1000, permissions: { maxLinesPerWrite: Infinity, maxFilesPerTask: Infinity, allowRiskPath: true, allowSubagentSpawn: true } },
-];
-export function getTierDefinition(tier) {
-    return TIER_DEFINITIONS[tier - 1];
-}
-export function getTierByPoints(totalPoints) {
-    // 从高到低检查，找到最高匹配等级
-    for (let i = TIER_DEFINITIONS.length - 1; i >= 0; i--) {
-        if (totalPoints >= TIER_DEFINITIONS[i].requiredPoints) {
-            return TIER_DEFINITIONS[i].tier;
-        }
-    }
-    return EvolutionTier.Seed;
-}
-export const TASK_DIFFICULTY_CONFIG = {
-    trivial: { basePoints: 1, description: '简单任务：读取、搜索、状态查询' },
-    normal: { basePoints: 3, description: '常规任务：单文件编辑、测试编写' },
-    hard: { basePoints: 8, description: '困难任务：多文件重构、架构变更' },
-};
-export const DEFAULT_EVOLUTION_CONFIG = {
-    doubleRewardCooldownMs: 60 * 60 * 1000, // 1小时
-    maxRecentEvents: 50,
-    difficultyPenalty: {
-        tier4Trivial: 0.1,
-        tier4Normal: 0.5,
-        tier5Trivial: 0.1,
-        tier5Normal: 0.5,
-    },
-};
-/**
- * Validates that a detector metadata object has all required fields with non-empty values.
- * Used as defense-in-depth before accepting detectorMetadata for auto-trainable principles.
- */
-export function isCompleteDetectorMetadata(meta) {
-    if (!meta || typeof meta !== 'object')
-        return false;
-    const m = meta;
-    const VALID_CONFIDENCE = ['high', 'medium', 'low'];
-    if (typeof m.confidence !== 'string' ||
-        !VALID_CONFIDENCE.includes(m.confidence)) {
-        return false;
-    }
-    const nonEmptyStringArray = (arr) => Array.isArray(arr) &&
-        arr.length > 0 &&
-        arr.every((s) => typeof s === 'string' && s.length > 0);
-    return (nonEmptyStringArray(m.applicabilityTags) &&
-        nonEmptyStringArray(m.positiveSignals) &&
-        nonEmptyStringArray(m.negativeSignals));
-}

package/dist/core/external-training-contract.d.ts

@@ -1,276 +0,0 @@
-/**
- * External Training Contract — Normalized Experiment Spec and Result Schema
- * ========================================================================
- *
- * PURPOSE: Define the stable contract between the plugin and external trainer
- * backends. The plugin produces a constrained experiment specification that an
- * external trainer consumes. The trainer returns a normalized result that the
- * plugin can register, evaluate, and gate for rollout.
- *
- * ARCHITECTURE:
- *   - Plugin is responsible for creating the experiment spec
- *   - Plugin is responsible for validating the trainer result
- *   - Plugin is responsible for registering lineage (train run → checkpoint → eval)
- *   - Plugin is responsible for invoking benchmark evaluation
- *   - Plugin is responsible for invoking promotion gate logic
- *   - Plugin is responsible for binding deployment only after gate approval
- *
- * DESIGN CONSTRAINTS:
- *   - ORPO-first: trainingMode must be 'orpo' for production runs
- *   - No real training inside the plugin
- *   - No direct deployment promotion from trainer output
- *   - No direct trainer writes to review/eval/deployment state
- *   - Backend-pluggable: same contract works for all backends
- *
- * CONTRACT GOALS:
- *   - support ORPO training for approved nocturnal exports
- *   - support multiple backend implementations behind one schema
- *   - preserve dataset / config / checkpoint lineage
- *   - remain valid on consumer hardware
- *   - fail closed when inputs are incomplete or inconsistent
- */
-/**
- * Allowed backend identifiers.
- *
- * - `peft-trl-orpo`: primary reference implementation using PEFT + TRL ORPO
- * - `unsloth-orpo`: compatible accelerated implementation using Unsloth
- * - `dry-run`: validates paths/spec/environment only, no real training
- */
-export type TrainerBackendKind = 'peft-trl-orpo' | 'unsloth-orpo' | 'dry-run';
-/**
- * Hardware tier for training.
- *
- * - `consumer-gpu`: RTX 4090 24GB or equivalent (production target)
- * - `small-gpu`: 8GB-16GB VRAM (compatibility target)
- * - `cpu-experimental`: CPU-only experimental runs (dry-run or tiny models only)
- */
-export type HardwareTier = 'consumer-gpu' | 'small-gpu' | 'cpu-experimental';
-/**
- * Worker profiles supported for training.
- *
- * Phase 7 first rollout: `local-reader` only.
- * `local-editor` requires explicit human approval to enable.
- */
-export type TrainableWorkerProfile = 'local-reader' | 'local-editor';
-/**
- * Training mode — Phase 7 production is ORPO-only.
- */
-export type TrainingMode = 'orpo';
-/**
- * Hyperparameters for ORPO training.
- */
-export interface TrainingHyperparameters {
-    learningRate: number;
-    batchSize: number;
-    gradientAccumulation: number;
-    loraRank: number;
-    loraAlpha: number;
-    loraDropout: number;
-    warmupRatio: number;
-    maxSteps: number;
-    maxSeqLength: number;
-}
-/**
- * Budget constraints for a training experiment.
- */
-export interface TrainingBudget {
-    maxWallClockMinutes: number;
-    maxTrainTokens?: number;
-}
-/**
- * Expected artifact from a successful training run.
- */
-export interface ExpectedArtifact {
-    checkpointName: string;
-    adapterFormat: 'peft-adapter';
-}
-/**
- * The experiment specification sent to an external trainer.
- * This defines WHAT to train, not HOW to train (backend-specific).
- */
-export interface TrainingExperimentSpec {
-    /** Unique identifier for this experiment */
-    experimentId: string;
-    /** Which backend to use */
-    backend: TrainerBackendKind;
-    /** Training mode — only 'orpo' is supported in Phase 7 */
-    trainingMode: TrainingMode;
-    /** Target worker profile for this experiment */
-    targetWorkerProfile: TrainableWorkerProfile;
-    /** Target model family to train */
-    targetModelFamily: string;
-    /** Hardware tier for this experiment */
-    hardwareTier: HardwareTier;
-    /** Reference to the ORPO export providing training data */
-    datasetExportId: string;
-    datasetExportPath: string;
-    /** Fingerprint of the dataset for lineage verification */
-    datasetFingerprint: string;
-    /** Reference to the benchmark export for eval */
-    benchmarkExportId: string;
-    /** Output directory for checkpoint artifacts */
-    outputDir: string;
-    /** Fingerprint of the training configuration */
-    configFingerprint: string;
-    /** Hash of the training code/contract version */
-    codeHash: string;
-    /** Training hyperparameters */
-    hyperparameters: TrainingHyperparameters;
-    /** Budget constraints */
-    budget: TrainingBudget;
-    /** Expected artifact from training */
-    expectedArtifact: ExpectedArtifact;
-}
-/**
- * Training metrics recorded by the backend.
- */
-export interface TrainingMetrics {
-    wallClockMinutes: number;
-    finalLoss?: number;
-    tokensSeen?: number;
-}
-/**
- * Artifact produced by a successful training run.
- */
-export interface TrainingArtifact {
-    adapterFormat: 'peft-adapter';
-    artifactPath: string;
-}
-/**
- * Status of a training experiment.
- */
-export type ExperimentStatus = 'completed' | 'failed' | 'dry_run';
-/**
- * The result returned by an external trainer after execution.
- * This defines the output contract — all backends must return the same shape.
- */
-export interface TrainingExperimentResult {
-    /** Experiment ID (must match the spec's experimentId) */
-    experimentId: string;
-    /** Which backend was used */
-    backend: TrainerBackendKind;
-    /** Final status of the experiment */
-    status: ExperimentStatus;
-    /** Registered training run ID (plugin-side) */
-    trainRunId?: string;
-    /** Registered checkpoint ID (plugin-side) */
-    checkpointId?: string;
-    /** Checkpoint reference string (for lineage) */
-    checkpointRef?: string;
-    /** Target worker profile */
-    targetWorkerProfile: TrainableWorkerProfile;
-    /** Target model family */
-    targetModelFamily: string;
-    /** Dataset fingerprint (for lineage verification) */
-    datasetFingerprint: string;
-    /** Config fingerprint (for lineage verification) */
-    configFingerprint: string;
-    /** Code hash (for lineage verification) */
-    codeHash: string;
-    /** Training metrics */
-    metrics?: TrainingMetrics;
-    /** Produced artifact (only if status === 'completed') */
-    artifact?: TrainingArtifact;
-    /** Failure reason (only if status === 'failed') */
-    failureReason?: string;
-    /** ISO-8601 creation timestamp */
-    createdAt: string;
-}
-/**
- * Validation error for trainer result verification.
- */
-export interface ValidationError {
-    field: string;
-    expected: string;
-    actual: string;
-    reason: string;
-}
-/**
- * Result of validating a trainer result against the experiment spec.
- */
-export interface ValidationResult {
-    valid: boolean;
-    errors: ValidationError[];
-}
-/**
- * Validate that a trainer result matches the experiment spec.
- *
- * FAILS CLOSED on any mismatch — a checkpoint with invalid lineage must not
- * be registered or promoted.
- *
- * Validation rules:
- * 1. experimentId must match
- * 2. backend must match
- * 3. targetWorkerProfile must match
- * 4. targetModelFamily must match
- * 5. datasetFingerprint must match
- * 6. configFingerprint must match
- * 7. codeHash must match
- * 8. dry-run must not produce a deployable checkpoint
- *
- * @param spec - The original experiment spec
- * @param result - The trainer result to validate
- * @returns ValidationResult indicating pass/fail and any errors
- */
-export declare function validateTrainerResult(spec: TrainingExperimentSpec, result: TrainingExperimentResult): ValidationResult;
-/**
- * Generate a fingerprint for a configuration object.
- * Used for configFingerprint in the experiment spec.
- */
-export declare function computeConfigFingerprint(config: Partial<TrainingHyperparameters>): string;
-/**
- * Generate a fingerprint for a dataset export.
- * Used for datasetFingerprint in the experiment spec.
- *
- * Combines file content hash with sampleCount to detect:
- * - Content changes (file modified/replaced)
- * - Sample count changes (different export)
- *
- * If the file cannot be read, falls back to path+count hash (legacy behavior).
- */
-export declare function computeDatasetFingerprint(exportPath: string, sampleCount: number): string;
-/**
- * Generate a code hash for the training contract version.
- * Used for codeHash in the experiment spec.
- *
- * Hashes the actual contract source file content so any change to the
- * contract produces a different hash, ensuring lineage integrity.
- *
- * Falls back to version string + timestamp if source cannot be read.
- */
-export declare function computeCodeHash(): string;
-/**
- * Generate a new experiment ID.
- */
-export declare function generateExperimentId(): string;
-/**
- * Validate that a hardware tier is appropriate for the backend.
- *
- * @param backend - The backend being used
- * @param tier - The hardware tier
- * @throws Error if the combination is not supported
- */
-export declare function validateHardwareTier(backend: TrainerBackendKind, tier: HardwareTier): void;
-/**
- * Get the default hardware tier for a backend.
- */
-export declare function getDefaultHardwareTier(backend: TrainerBackendKind): HardwareTier;
-/**
- * Valid model family patterns for local-reader profile.
- * Used for family validation in the training contract.
- */
-export declare const READER_FAMILY_PATTERNS: string[];
-/**
- * Valid model family patterns for local-editor profile.
- * Used for family validation in the training contract.
- */
-export declare const EDITOR_FAMILY_PATTERNS: string[];
-/**
- * Check if a model family is valid for a worker profile.
- */
-export declare function isValidModelFamilyForProfile(family: string, profile: TrainableWorkerProfile): boolean;
-/**
- * Phase 7 first rollout is limited to local-reader.
- * This flag controls whether local-editor is allowed.
- */
-export declare const LOCAL_EDITOR_ENABLED = false;