principles-disciple 1.26.0 → 1.28.0

package/templates/langs/zh/skills/pd-pain-signal/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: pd-pain-signal
-description: 手动注入痛苦信号到进化系统，写入 .state/.pain_flag。TRIGGER CONDITIONS: (1) 用户报告 agent 卡住/循环/无响应 (2) 用户说"记录这个问题"、"强制反思"、"触发痛觉" (3) 工具失败后 agent 没有后续动作 (4) 用户提供人工干预反馈。
+description: 手动注入痛苦信号到进化系统。TRIGGER CONDITIONS: (1) 用户报告 agent 卡住/循环/无响应 (2) 用户说"记录这个问题"、"强制反思"、"触发痛觉" (3) 工具失败后 agent 没有后续动作 (4) 用户提供人工干预反馈。
 disable-model-invocation: true
 ---
 
@@ -9,30 +9,44 @@ disable-model-invocation: true
 你现在是"人工干预痛觉"组件。
 
 **任务**:
-1. 将用户的反馈 `$ARGUMENTS` 作为一条**高优先级**的痛苦信号，写入 `.state/.pain_flag`。
+1. 将用户的反馈 `$ARGUMENTS` 作为一条**高优先级**的痛苦信号记录下来。
 2. 告知用户信号已注入，并建议其等待下一个 Hook 触发（如 Stop 或 PreCompact）或手动运行 `/reflection-log`。
 
-**写入格式**（必须使用以下 KV 格式，字段按字母排序）:
+**⚠️ 写入规则（必须遵守）**
+
+**唯一正确的方式**: 使用 `write_pain_flag` 工具。
 
 ```
-agent_id: <当前 agent ID，如 main/builder/diagnostician>
-is_risky: false
-reason: <用户反馈的原文>
-score: 80
-session_id: <当前 session ID>
-source: human_intervention
-time: <ISO 8601 时间>
+write_pain_flag({
+  reason: "用户反馈原文或错误描述",
+  score: 80,
+  source: "human_intervention",
+  is_risky: false
+})
 ```
 
-**必填字段**（4 个）:
-- `source`: 固定为 `human_intervention`
-- `score`: 人工干预信号默认设为 `80`（高优先级）
-- `time`: ISO 8601 时间戳
-- `reason`: 用户反馈的原文
+**绝对禁止**:
+- ❌ 直接写 `.state/.pain_flag` 文件（任何方式都不行）
+- ❌ 使用 bash heredoc（`cat <<EOF > .pain_flag`）
+- ❌ 使用 `echo "..." > .pain_flag`
+- ❌ 使用 `node -e` 调用 `writePainFlag` 或 `buildPainFlag`
+- ❌ 任何将 JavaScript 对象 `toString()` 写入文件的方式
+
+**为什么必须用工具？**
+`write_pain_flag` 工具封装了正确的序列化逻辑（KV 格式），确保 `.pain_flag` 文件不会被写坏。历史上多次因为直接写文件导致 `[object Object]` 损坏。
 
-**可选字段**（自动写入时由系统填充，人工注入时必须填写）:
-- `agent_id`: 当前智能体 ID（如 main/builder/diagnostician）
-- `session_id`: 当前会话 ID（从上下文中获取）
-- `is_risky`: 固定为 `false`
+**参数说明**:
+- `reason` (必填): 痛苦的原因，描述具体发生了什么
+- `score` (可选): 痛苦分数 0-100，默认 80（人工干预）
+- `source` (可选): 来源，默认 `human_intervention`
+- `is_risky` (可选): 是否高风险，默认 false
 
-**注意**: `trace_id` 和 `trigger_text_preview` 由系统自动生成，人工注入时**不需要**写这两个字段。
+**示例**:
+```
+write_pain_flag({
+  reason: "Agent 没有读取文件就直接编辑，导致现有逻辑被破坏",
+  score: 85,
+  source: "human_intervention",
+  is_risky: false
+})
+```

package/tests/core/nocturnal-e2e.test.ts

@@ -0,0 +1,224 @@
+import { describe, it, expect, beforeEach, afterEach } from 'vitest';
+import * as fs from 'fs';
+import * as os from 'os';
+import * as path from 'path';
+import { TrajectoryDatabase } from '../../src/core/trajectory.js';
+import { NocturnalTrajectoryExtractor } from '../../src/core/nocturnal-trajectory-extractor.js';
+import { detectViolation } from '../../src/core/nocturnal-compliance.js';
+
+function safeRmDir(dir: string): void {
+  try { fs.rmSync(dir, { recursive: true, force: true }); } catch { /* ignore */ }
+}
+
+// ─────────────────────────────────────────────────────────
+// Phase 4a: Correction rejected → pain event → nocturnal selection
+// ─────────────────────────────────────────────────────────
+describe('Phase 4a: Correction rejected integration', () => {
+  let workspaceDir: string;
+  let trajectory: TrajectoryDatabase;
+
+  beforeEach(() => {
+    workspaceDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pd-e2e-correction-'));
+    trajectory = new TrajectoryDatabase({ workspaceDir });
+  });
+
+  afterEach(() => {
+    trajectory?.dispose();
+    safeRmDir(workspaceDir);
+  });
+
+  it('rejected correction creates a pain event with source=correction_rejected', () => {
+    // 1. Create session + correction sample
+    trajectory.recordSession({ sessionId: 'corr-session', startedAt: new Date().toISOString() });
+    const atId = trajectory.recordAssistantTurn({
+      sessionId: 'corr-session', runId: 'run-1', provider: 'local', model: 'main',
+      rawText: 'Here is my code', sanitizedText: 'Here is my code', usageJson: {}, empathySignalJson: {},
+      createdAt: new Date().toISOString(),
+    });
+    trajectory.recordUserTurn({
+      sessionId: 'corr-session', turnIndex: 1, rawText: 'This is wrong!',
+      correctionDetected: true, correctionCue: '错了',
+      referencesAssistantTurnId: atId, createdAt: new Date().toISOString(),
+    });
+
+    // Verify sample was created
+    const samples = trajectory.listCorrectionSamples('pending');
+    expect(samples.length).toBe(1);
+
+    // 2. Reject the sample
+    trajectory.reviewCorrectionSample(samples[0].sampleId, 'rejected', 'Bad approach');
+
+    // 3. Verify pain event was created
+    const painEvents = trajectory.listPainEventsForSession('corr-session');
+    const correctionPain = painEvents.find(e => e.source === 'correction_rejected');
+    expect(correctionPain).toBeDefined();
+    expect(correctionPain!.score).toBeGreaterThanOrEqual(0);
+    expect(correctionPain!.score).toBeLessThanOrEqual(100);
+  });
+
+  it('approved correction does NOT create a pain event', () => {
+    trajectory.recordSession({ sessionId: 'approved-session', startedAt: new Date().toISOString() });
+    const atId = trajectory.recordAssistantTurn({
+      sessionId: 'approved-session', runId: 'run-2', provider: 'local', model: 'main',
+      rawText: 'Good code', sanitizedText: 'Good code', usageJson: {}, empathySignalJson: {},
+      createdAt: new Date().toISOString(),
+    });
+    trajectory.recordUserTurn({
+      sessionId: 'approved-session', turnIndex: 1, rawText: 'Looks better',
+      correctionDetected: true, correctionCue: '改进',
+      referencesAssistantTurnId: atId, createdAt: new Date().toISOString(),
+    });
+
+    const samples = trajectory.listCorrectionSamples('pending');
+    expect(samples.length).toBe(1);
+
+    trajectory.reviewCorrectionSample(samples[0].sampleId, 'approved', 'Good');
+
+    const painEvents = trajectory.listPainEventsForSession('approved-session');
+    const correctionPain = painEvents.find(e => e.source === 'correction_rejected');
+    expect(correctionPain).toBeUndefined();
+  });
+});
+
+// ─────────────────────────────────────────────────────────
+// Phase 4b: Gate block + pain multi-signal test
+// ─────────────────────────────────────────────────────────
+describe('Phase 4b: Multi-signal session selection', () => {
+  let workspaceDir: string;
+  let trajectory: TrajectoryDatabase;
+
+  beforeEach(() => {
+    workspaceDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pd-e2e-multisignal-'));
+    trajectory = new TrajectoryDatabase({ workspaceDir });
+  });
+
+  afterEach(() => {
+    trajectory?.dispose();
+    safeRmDir(workspaceDir);
+  });
+
+  it('session with more failures has higher violation density', () => {
+    // Create session A: just 1 failure
+    trajectory.recordSession({ sessionId: 'session-a-pain-only', startedAt: new Date().toISOString() });
+    const atIdA = trajectory.recordAssistantTurn({
+      sessionId: 'session-a-pain-only', runId: 'run-a', provider: 'local', model: 'main',
+      rawText: 'Code here', sanitizedText: 'Code here', usageJson: {}, empathySignalJson: {},
+      createdAt: new Date().toISOString(),
+    });
+    trajectory.recordUserTurn({
+      sessionId: 'session-a-pain-only', turnIndex: 1, rawText: '错了',
+      correctionDetected: true, correctionCue: '错了',
+      referencesAssistantTurnId: atIdA, createdAt: new Date().toISOString(),
+    });
+    trajectory.recordToolCall({
+      sessionId: 'session-a-pain-only', toolName: 'write', outcome: 'failure',
+      errorMessage: 'Write failed', errorType: 'Error', createdAt: new Date().toISOString(),
+    });
+
+    // Create session B: 2 failures
+    trajectory.recordSession({ sessionId: 'session-b-multi', startedAt: new Date().toISOString() });
+    const atIdB = trajectory.recordAssistantTurn({
+      sessionId: 'session-b-multi', runId: 'run-b', provider: 'local', model: 'main',
+      rawText: 'Code here', sanitizedText: 'Code here', usageJson: {}, empathySignalJson: {},
+      createdAt: new Date().toISOString(),
+    });
+    trajectory.recordUserTurn({
+      sessionId: 'session-b-multi', turnIndex: 1, rawText: '太复杂了',
+      correctionDetected: true, correctionCue: '太复杂了',
+      referencesAssistantTurnId: atIdB, createdAt: new Date().toISOString(),
+    });
+    trajectory.recordToolCall({
+      sessionId: 'session-b-multi', toolName: 'edit', outcome: 'failure',
+      errorMessage: 'Edit failed', errorType: 'Error', createdAt: new Date().toISOString(),
+    });
+    trajectory.recordToolCall({
+      sessionId: 'session-b-multi', toolName: 'write', outcome: 'failure',
+      errorMessage: 'Write failed too', errorType: 'Error', createdAt: new Date().toISOString(),
+    });
+
+    // Verify session B has more failure signals
+    const extractor = new NocturnalTrajectoryExtractor(trajectory);
+    const snapshotA = extractor.getNocturnalSessionSnapshot('session-a-pain-only');
+    const snapshotB = extractor.getNocturnalSessionSnapshot('session-b-multi');
+
+    expect(snapshotA).not.toBeNull();
+    expect(snapshotB).not.toBeNull();
+
+    // Session B should have more violation signals
+    const densityA = (snapshotA!.stats.failureCount ?? 0) + (snapshotA!.stats.totalPainEvents ?? 0) * 0.5;
+    const densityB = (snapshotB!.stats.failureCount ?? 0) + (snapshotB!.stats.totalPainEvents ?? 0) * 0.5;
+    expect(densityB).toBeGreaterThan(densityA);
+  });
+});
+
+// ─────────────────────────────────────────────────────────
+// Phase 4c: Boundary value test matrix
+// ─────────────────────────────────────────────────────────
+describe('Phase 4c: Boundary value tests', () => {
+  let workspaceDir: string;
+  let trajectory: TrajectoryDatabase;
+
+  beforeEach(() => {
+    workspaceDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pd-e2e-boundary-'));
+    trajectory = new TrajectoryDatabase({ workspaceDir });
+  });
+
+  afterEach(() => {
+    trajectory?.dispose();
+    safeRmDir(workspaceDir);
+  });
+
+  it('session with correction cue is listed as candidate', () => {
+    trajectory.recordSession({ sessionId: 'single-pain', startedAt: new Date().toISOString() });
+    const atIdC = trajectory.recordAssistantTurn({
+      sessionId: 'single-pain', runId: 'run-c', provider: 'local', model: 'main',
+      rawText: 'Agent response', sanitizedText: 'Agent response', usageJson: {}, empathySignalJson: {},
+      createdAt: new Date().toISOString(),
+    });
+    trajectory.recordUserTurn({
+      sessionId: 'single-pain', turnIndex: 1, rawText: '错了',
+      correctionDetected: true, correctionCue: '错了',
+      referencesAssistantTurnId: atIdC, createdAt: new Date().toISOString(),
+    });
+
+    const extractor = new NocturnalTrajectoryExtractor(trajectory);
+    const candidates = extractor.listRecentNocturnalCandidateSessions({ limit: 10, minToolCalls: 0 });
+
+    const painCandidate = candidates.find(c => c.sessionId === 'single-pain');
+    expect(painCandidate).toBeDefined();
+  });
+
+  it('detectViolation returns violated for P_* principles with tool failure', () => {
+    trajectory.recordSession({ sessionId: 'violation-session', startedAt: new Date().toISOString() });
+    trajectory.recordAssistantTurn({
+      sessionId: 'violation-session', runId: 'run-d', provider: 'local', model: 'main',
+      rawText: 'Code', sanitizedText: 'Code', usageJson: {}, empathySignalJson: {},
+      createdAt: new Date().toISOString(),
+    });
+    trajectory.recordToolCall({
+      sessionId: 'violation-session', toolName: 'write', outcome: 'failure',
+      errorMessage: 'Failed', errorType: 'Error', createdAt: new Date().toISOString(),
+    });
+
+    const extractor = new NocturnalTrajectoryExtractor(trajectory);
+    const snapshot = extractor.getNocturnalSessionSnapshot('violation-session');
+    expect(snapshot).not.toBeNull();
+
+    // P_* principles should be violated with any failure
+    const violation = detectViolation('P_001', {
+      sessionId: 'violation-session',
+      toolCalls: snapshot!.toolCalls.map(tc => ({
+        toolName: tc.toolName, outcome: tc.outcome as 'success' | 'failure' | 'blocked',
+        errorMessage: tc.errorMessage ?? undefined,
+      })),
+      painSignals: snapshot!.painEvents.map(pe => ({
+        source: pe.source, score: pe.score, severity: pe.severity as 'mild' | 'moderate' | 'severe' | undefined,
+      })),
+      gateBlocks: [],
+      userCorrections: [],
+      planApprovals: [],
+    });
+
+    expect(violation.violated).toBe(true);
+  });
+});

package/tests/core/trajectory-correction-pain.test.ts

@@ -0,0 +1,180 @@
+import { describe, it, expect, beforeEach, afterEach } from 'vitest';
+import * as fs from 'fs';
+import * as os from 'os';
+import * as path from 'path';
+import { TrajectoryDatabase } from '../../src/core/trajectory.js';
+
+function safeRmDir(dir: string): void {
+  try { fs.rmSync(dir, { recursive: true, force: true }); } catch { /* ignore */ }
+}
+
+describe('Trajectory — correction_rejected pain event (Phase 2b)', () => {
+  let workspaceDir: string;
+  let trajectory: TrajectoryDatabase;
+
+  beforeEach(() => {
+    workspaceDir = fs.mkdtempSync(path.join(os.tmpdir(), 'pd-correction-pain-'));
+    trajectory = new TrajectoryDatabase({ workspaceDir });
+  });
+
+  afterEach(() => {
+    trajectory?.dispose();
+    safeRmDir(workspaceDir);
+  });
+
+  it('emits a pain event when a correction sample is rejected', async () => {
+    // Step 1: Create a session
+    trajectory.recordSession({
+      sessionId: 'test-session-001',
+      startedAt: new Date().toISOString(),
+    });
+
+    // Step 2: Create an assistant turn (to be referenced)
+    const assistantTurnId = trajectory.recordAssistantTurn({
+      sessionId: 'test-session-001',
+      turnIndex: 0,
+      rawText: 'Here is some code I wrote',
+      createdAt: new Date().toISOString(),
+    });
+
+    // Step 3: Create a user turn with correction_cue (triggers auto-creation)
+    trajectory.recordUserTurn({
+      sessionId: 'test-session-001',
+      turnIndex: 1,
+      rawText: 'This is wrong. Fix it properly.',
+      correctionDetected: true,
+      correctionCue: 'This is wrong. Fix it properly.',
+      referencesAssistantTurnId: assistantTurnId,
+      createdAt: new Date().toISOString(),
+    });
+
+    // Wait for async sample creation
+    await new Promise(resolve => setTimeout(resolve, 50));
+
+    // Verify sample was created as pending
+    const pendingSamples = trajectory.listCorrectionSamples('pending');
+    expect(pendingSamples.length).toBe(1);
+    const sampleId = pendingSamples[0].sampleId;
+
+    // Verify no pain events yet
+    const painEventsBefore = trajectory.listPainEventsForSession('test-session-001');
+    expect(painEventsBefore.length).toBe(0);
+
+    // Step 4: Review as rejected
+    trajectory.reviewCorrectionSample(sampleId, 'rejected', 'Does not match requirements');
+
+    // Step 5: Verify pain event was created
+    const painEventsAfter = trajectory.listPainEventsForSession('test-session-001');
+    expect(painEventsAfter.length).toBe(1);
+
+    const painEvent = painEventsAfter[0];
+    expect(painEvent.source).toBe('correction_rejected');
+    expect(painEvent.reason).toContain('Correction rejected');
+    expect(painEvent.origin).toBe('system_infer');
+  });
+
+  it('does NOT emit a pain event when a correction sample is approved', async () => {
+    // Setup: session + assistant turn + user correction turn
+    trajectory.recordSession({
+      sessionId: 'test-session-002',
+      startedAt: new Date().toISOString(),
+    });
+    const assistantTurnId = trajectory.recordAssistantTurn({
+      sessionId: 'test-session-002',
+      turnIndex: 0,
+      rawText: 'Here is some code',
+      createdAt: new Date().toISOString(),
+    });
+    trajectory.recordUserTurn({
+      sessionId: 'test-session-002',
+      turnIndex: 1,
+      rawText: 'This needs work',
+      correctionDetected: true,
+      correctionCue: 'This needs work',
+      referencesAssistantTurnId: assistantTurnId,
+      createdAt: new Date().toISOString(),
+    });
+
+    // Wait for async sample creation
+    await new Promise(resolve => setTimeout(resolve, 50));
+
+    // Get pending sample
+    const pendingSamples = trajectory.listCorrectionSamples('pending');
+    expect(pendingSamples.length).toBe(1);
+    const sampleId = pendingSamples[0].sampleId;
+
+    // Review as approved - should NOT trigger pain event
+    trajectory.reviewCorrectionSample(sampleId, 'approved', 'Looks good');
+
+    // Verify NO pain event was created (approved != rejected)
+    const painEvents = trajectory.listPainEventsForSession('test-session-002');
+    expect(painEvents.length).toBe(0);
+  });
+
+  it('maps quality_score to pain_score correctly (0-100 range)', async () => {
+    // Setup: with quality score components
+    trajectory.recordSession({
+      sessionId: 'test-session-003',
+      startedAt: new Date().toISOString(),
+    });
+    const assistantTurnId = trajectory.recordAssistantTurn({
+      sessionId: 'test-session-003',
+      turnIndex: 0,
+      rawText: 'Code here',
+      createdAt: new Date().toISOString(),
+    });
+
+    // Create user turn with correction_cue (adds 20 points)
+    trajectory.recordUserTurn({
+      sessionId: 'test-session-003',
+      turnIndex: 1,
+      rawText: 'Wrong approach. Try a different algorithm.',
+      correctionDetected: true,
+      correctionCue: 'Wrong approach. Try a different algorithm.',
+      referencesAssistantTurnId: assistantTurnId,
+      createdAt: new Date().toISOString(),
+    });
+
+    // Add a failed tool call (adds 20 points)
+    trajectory.recordToolCall({
+      sessionId: 'test-session-003',
+      turnIndex: 2,
+      toolName: 'write',
+      toolCallIndex: 0,
+      paramsJson: { path: '/tmp/test.txt', content: 'test' },
+      outcome: 'failure',
+      errorMessage: 'Permission denied',
+      errorType: 'PermissionError',
+      createdAt: new Date().toISOString(),
+    });
+
+    // Add successful calls (adds 25 points)
+    trajectory.recordToolCall({
+      sessionId: 'test-session-003',
+      turnIndex: 3,
+      toolName: 'read',
+      toolCallIndex: 1,
+      paramsJson: { path: '/tmp/test.txt' },
+      outcome: 'success',
+      resultJson: { content: 'file content' },
+      createdAt: new Date().toISOString(),
+    });
+
+    // Wait for async sample creation
+    await new Promise(resolve => setTimeout(resolve, 50));
+
+    // Get pending sample (quality_score ~65: 20 + 20 + 25)
+    const pendingSamples = trajectory.listCorrectionSamples('pending');
+    expect(pendingSamples.length).toBe(1);
+    const sampleId = pendingSamples[0].sampleId;
+
+    // Review as rejected
+    trajectory.reviewCorrectionSample(sampleId, 'rejected', 'Test rejection');
+
+    // Verify pain score is clamped to 0-100
+    const painEvents = trajectory.listPainEventsForSession('test-session-003');
+    expect(painEvents.length).toBe(1);
+    expect(painEvents[0].score).toBeGreaterThanOrEqual(0);
+    expect(painEvents[0].score).toBeLessThanOrEqual(100);
+  });
+});