principles-disciple 1.28.0 → 1.28.1
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/openclaw.plugin.json +1 -1
- package/package.json +4 -4
- package/scripts/validate-live-path.ts +18 -18
- package/src/commands/nocturnal-train.ts +4 -6
- package/src/commands/pain.ts +8 -11
- package/src/commands/pd-reflect.ts +1 -1
- package/src/core/bootstrap-rules.ts +3 -3
- package/src/core/merge-gate-audit.ts +1 -1
- package/src/core/nocturnal-candidate-scoring.ts +131 -0
- package/src/core/nocturnal-reasoning-deriver.ts +337 -0
- package/src/core/nocturnal-trinity.ts +454 -18
- package/src/core/pain-context-extractor.ts +1 -3
- package/src/core/principle-tree-migration.ts +2 -4
- package/src/core/thinking-os-parser.ts +3 -3
- package/src/hooks/bash-risk.ts +1 -1
- package/src/hooks/gfi-gate.ts +1 -1
- package/src/hooks/pain.ts +1 -1
- package/src/hooks/prompt.ts +36 -2
- package/src/hooks/subagent.ts +1 -1
- package/src/index.ts +1 -1
- package/src/service/evolution-worker.ts +1 -1
- package/src/service/health-query-service.ts +15 -6
- package/src/service/subagent-workflow/nocturnal-workflow-manager.ts +0 -1
- package/tests/core/nocturnal-candidate-scoring.test.ts +132 -0
- package/tests/core/nocturnal-reasoning-deriver.test.ts +372 -0
- package/tests/core/nocturnal-trinity.test.ts +791 -0
|
@@ -0,0 +1,337 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Nocturnal Reasoning Deriver — Runtime Reasoning Signal Extraction
|
|
3
|
+
* ==============================================================
|
|
4
|
+
*
|
|
5
|
+
* PURPOSE: Derive structured reasoning signals from existing snapshot data
|
|
6
|
+
* without any snapshot schema changes. Pure functions, zero dependencies.
|
|
7
|
+
*
|
|
8
|
+
* THREE FUNCTIONS:
|
|
9
|
+
* - deriveReasoningChain: Extract thinking content, uncertainty, confidence from assistant turns
|
|
10
|
+
* - deriveDecisionPoints: Extract before/after context per tool call (Plan 02)
|
|
11
|
+
* - deriveContextualFactors: Compute contextual factors from snapshot (Plan 02)
|
|
12
|
+
*/
|
|
13
|
+
|
|
14
|
+
import type { NocturnalAssistantTurn, NocturnalToolCall, NocturnalUserTurn, NocturnalSessionSnapshot } from './nocturnal-trajectory-extractor.js';
|
|
15
|
+
import { detectThinkingModelMatches, listThinkingModels } from './thinking-models.js';
|
|
16
|
+
|
|
17
|
+
// ---------------------------------------------------------------------------
|
|
18
|
+
// Shared helpers
|
|
19
|
+
// ---------------------------------------------------------------------------
|
|
20
|
+
|
|
21
|
+
/** Parse an ISO 8601 timestamp, returning NaN for invalid formats. */
|
|
22
|
+
function parseTs(ts: string): number {
|
|
23
|
+
// ISO 8601 strings without Z suffix or offset are treated as local time.
|
|
24
|
+
// Log a warning for ambiguous formats (missing timezone indicator).
|
|
25
|
+
if (
|
|
26
|
+
typeof ts === 'string' &&
|
|
27
|
+
!ts.endsWith('Z') &&
|
|
28
|
+
!ts.includes('+') &&
|
|
29
|
+
ts.includes('-', 4)
|
|
30
|
+
) {
|
|
31
|
+
// Looks like an ISO date but no timezone — could be ambiguous
|
|
32
|
+
const bare = ts.slice(0, 10);
|
|
33
|
+
if (/^\d{4}-\d{2}-\d{2}$/.test(bare)) {
|
|
34
|
+
// Plain YYYY-MM-DD without time or Z — definitely ambiguous
|
|
35
|
+
console.warn(`[Deriver] Timestamp missing timezone: "${ts}"`);
|
|
36
|
+
}
|
|
37
|
+
}
|
|
38
|
+
return Date.parse(ts);
|
|
39
|
+
}
|
|
40
|
+
|
|
41
|
+
// ---------------------------------------------------------------------------
|
|
42
|
+
// Shared types (used across all three derive functions)
|
|
43
|
+
// ---------------------------------------------------------------------------
|
|
44
|
+
|
|
45
|
+
export interface DerivedReasoningSignal {
|
|
46
|
+
turnIndex: number;
|
|
47
|
+
thinkingContent: string;
|
|
48
|
+
uncertaintyMarkers: string[];
|
|
49
|
+
confidenceSignal: "high" | "medium" | "low";
|
|
50
|
+
}
|
|
51
|
+
|
|
52
|
+
export interface DerivedDecisionPoint {
|
|
53
|
+
toolName: string;
|
|
54
|
+
outcome: "success" | "failure" | "blocked";
|
|
55
|
+
beforeContext: string;
|
|
56
|
+
afterReflection?: string;
|
|
57
|
+
confidenceDelta?: number;
|
|
58
|
+
}
|
|
59
|
+
|
|
60
|
+
export interface DerivedContextualFactors {
|
|
61
|
+
fileStructureKnown: boolean;
|
|
62
|
+
errorHistoryPresent: boolean;
|
|
63
|
+
userGuidanceAvailable: boolean;
|
|
64
|
+
timePressure: boolean;
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
// ---------------------------------------------------------------------------
|
|
68
|
+
// Constants
|
|
69
|
+
// ---------------------------------------------------------------------------
|
|
70
|
+
|
|
71
|
+
const UNCERTAINTY_PATTERNS: RegExp[] = [
|
|
72
|
+
/let me (check|verify|confirm|understand)/gi,
|
|
73
|
+
/I should (first|probably|maybe)/gi,
|
|
74
|
+
/not sure (if|whether|about)/gi,
|
|
75
|
+
];
|
|
76
|
+
|
|
77
|
+
const THINKING_TAG_REGEX = /<thinking>([\s\S]*?)<\/thinking>/g;
|
|
78
|
+
|
|
79
|
+
// ---------------------------------------------------------------------------
|
|
80
|
+
// Helpers
|
|
81
|
+
// ---------------------------------------------------------------------------
|
|
82
|
+
|
|
83
|
+
/**
|
|
84
|
+
* Compute thinking model activation ratio for text.
|
|
85
|
+
* Uses detectThinkingModelMatches() from thinking-models.ts.
|
|
86
|
+
* Returns 0-1, rounded to 2 decimal places.
|
|
87
|
+
*/
|
|
88
|
+
function computeThinkingModelActivation(text: string): number {
|
|
89
|
+
if (!text || text.trim().length === 0) return 0;
|
|
90
|
+
const matches = detectThinkingModelMatches(text);
|
|
91
|
+
const totalModels = listThinkingModels().length;
|
|
92
|
+
if (totalModels === 0) return 0;
|
|
93
|
+
return Math.round((matches.length / totalModels) * 100) / 100;
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
/**
|
|
97
|
+
* Map activation ratio (0-1) to confidence signal.
|
|
98
|
+
* Thresholds: high > 0.6, medium 0.3-0.6, low < 0.3
|
|
99
|
+
*/
|
|
100
|
+
function mapConfidenceSignal(activation: number): "high" | "medium" | "low" {
|
|
101
|
+
if (activation > 0.6) return "high";
|
|
102
|
+
if (activation >= 0.3) return "medium";
|
|
103
|
+
return "low";
|
|
104
|
+
}
|
|
105
|
+
|
|
106
|
+
// ---------------------------------------------------------------------------
|
|
107
|
+
// deriveReasoningChain (DERIV-01)
|
|
108
|
+
// ---------------------------------------------------------------------------
|
|
109
|
+
|
|
110
|
+
/**
|
|
111
|
+
* Extract thinking content, uncertainty markers, and confidence signal
|
|
112
|
+
* from each assistant turn in the snapshot.
|
|
113
|
+
*
|
|
114
|
+
* DERIV-01: Returns one DerivedReasoningSignal per assistant turn.
|
|
115
|
+
* Empty input returns empty array. Never throws.
|
|
116
|
+
*/
|
|
117
|
+
export function deriveReasoningChain(assistantTurns: NocturnalAssistantTurn[]): DerivedReasoningSignal[] {
|
|
118
|
+
if (!assistantTurns || assistantTurns.length === 0) return [];
|
|
119
|
+
|
|
120
|
+
return assistantTurns.map(turn => {
|
|
121
|
+
const text = turn.sanitizedText ?? '';
|
|
122
|
+
|
|
123
|
+
// Extract all <thinking> content blocks (multiple blocks per turn possible)
|
|
124
|
+
const thinkingMatches = [...text.matchAll(THINKING_TAG_REGEX)];
|
|
125
|
+
const thinkingContent = thinkingMatches.map(m => m[1].trim()).join('\n');
|
|
126
|
+
|
|
127
|
+
// Detect uncertainty markers (collect all unique matches across 3 patterns)
|
|
128
|
+
const uncertaintyMarkers: string[] = [];
|
|
129
|
+
for (const pattern of UNCERTAINTY_PATTERNS) {
|
|
130
|
+
// Reset lastIndex to avoid g-flag state issues
|
|
131
|
+
pattern.lastIndex = 0;
|
|
132
|
+
const matches = text.match(pattern);
|
|
133
|
+
if (matches) {
|
|
134
|
+
for (const m of matches) {
|
|
135
|
+
if (!uncertaintyMarkers.includes(m)) {
|
|
136
|
+
uncertaintyMarkers.push(m);
|
|
137
|
+
}
|
|
138
|
+
}
|
|
139
|
+
}
|
|
140
|
+
}
|
|
141
|
+
|
|
142
|
+
// Confidence signal: only meaningful when <thinking> content exists.
|
|
143
|
+
// Without thinking tags we cannot extract a genuine reasoning trace, so
|
|
144
|
+
// we fall back to 'low' rather than misleading the downstream pipeline
|
|
145
|
+
// with activation derived from non-thinking patterns in the response text.
|
|
146
|
+
let confidenceSignal: "high" | "medium" | "low";
|
|
147
|
+
if (thinkingContent.length === 0) {
|
|
148
|
+
confidenceSignal = 'low';
|
|
149
|
+
} else {
|
|
150
|
+
const activation = computeThinkingModelActivation(text);
|
|
151
|
+
confidenceSignal = mapConfidenceSignal(activation);
|
|
152
|
+
}
|
|
153
|
+
|
|
154
|
+
return {
|
|
155
|
+
turnIndex: turn.turnIndex,
|
|
156
|
+
thinkingContent,
|
|
157
|
+
uncertaintyMarkers,
|
|
158
|
+
confidenceSignal,
|
|
159
|
+
};
|
|
160
|
+
});
|
|
161
|
+
}
|
|
162
|
+
|
|
163
|
+
// ---------------------------------------------------------------------------
|
|
164
|
+
// Helpers (Plan 02)
|
|
165
|
+
// ---------------------------------------------------------------------------
|
|
166
|
+
|
|
167
|
+
/**
|
|
168
|
+
* Convert confidence signal to numeric value for delta computation.
|
|
169
|
+
* high=1, medium=0.5, low=0
|
|
170
|
+
*/
|
|
171
|
+
function confidenceToNumber(signal: "high" | "medium" | "low"): number {
|
|
172
|
+
switch (signal) {
|
|
173
|
+
case "high": return 1;
|
|
174
|
+
case "medium": return 0.5;
|
|
175
|
+
case "low": return 0;
|
|
176
|
+
}
|
|
177
|
+
}
|
|
178
|
+
|
|
179
|
+
// ---------------------------------------------------------------------------
|
|
180
|
+
// deriveDecisionPoints (DERIV-02)
|
|
181
|
+
// ---------------------------------------------------------------------------
|
|
182
|
+
|
|
183
|
+
/**
|
|
184
|
+
* Extract before-context and after-reflection for each tool call.
|
|
185
|
+
*
|
|
186
|
+
* DERIV-02: For each tool call, find the assistant turn immediately before it
|
|
187
|
+
* (by createdAt timestamp) and extract last 500 chars as beforeContext.
|
|
188
|
+
* On failure outcome, find the next assistant turn and extract first 300 chars
|
|
189
|
+
* as afterReflection. Compute confidence delta between before/after.
|
|
190
|
+
*
|
|
191
|
+
* Empty inputs return empty array. Never throws.
|
|
192
|
+
*/
|
|
193
|
+
export function deriveDecisionPoints(
|
|
194
|
+
assistantTurns: NocturnalAssistantTurn[],
|
|
195
|
+
toolCalls: NocturnalToolCall[],
|
|
196
|
+
): DerivedDecisionPoint[] {
|
|
197
|
+
if (!toolCalls || toolCalls.length === 0) return [];
|
|
198
|
+
if (!assistantTurns || assistantTurns.length === 0) {
|
|
199
|
+
// Return decision points with empty beforeContext when no assistant turns
|
|
200
|
+
return toolCalls.map(tc => ({
|
|
201
|
+
toolName: tc.toolName,
|
|
202
|
+
outcome: tc.outcome,
|
|
203
|
+
beforeContext: '',
|
|
204
|
+
}));
|
|
205
|
+
}
|
|
206
|
+
|
|
207
|
+
// Sort assistant turns by createdAt for binary search
|
|
208
|
+
const sortedTurns = [...assistantTurns].sort(
|
|
209
|
+
(a, b) => parseTs(a.createdAt) - parseTs(b.createdAt)
|
|
210
|
+
);
|
|
211
|
+
|
|
212
|
+
// Binary search: find rightmost assistant turn with createdAt < tcTime
|
|
213
|
+
const findBeforeTurn = (tcTime: number): NocturnalAssistantTurn | undefined => {
|
|
214
|
+
let lo = 0, hi = sortedTurns.length - 1, result: NocturnalAssistantTurn | undefined;
|
|
215
|
+
while (lo <= hi) {
|
|
216
|
+
const mid = (lo + hi) >>> 1;
|
|
217
|
+
if (parseTs(sortedTurns[mid].createdAt) < tcTime) {
|
|
218
|
+
result = sortedTurns[mid];
|
|
219
|
+
lo = mid + 1;
|
|
220
|
+
} else {
|
|
221
|
+
hi = mid - 1;
|
|
222
|
+
}
|
|
223
|
+
}
|
|
224
|
+
return result;
|
|
225
|
+
};
|
|
226
|
+
|
|
227
|
+
return toolCalls.map(tc => {
|
|
228
|
+
const tcTime = parseTs(tc.createdAt);
|
|
229
|
+
const beforeTurn = findBeforeTurn(tcTime);
|
|
230
|
+
|
|
231
|
+
const beforeContext = beforeTurn
|
|
232
|
+
? beforeTurn.sanitizedText.slice(-500)
|
|
233
|
+
: '';
|
|
234
|
+
|
|
235
|
+
// On failure, find next assistant turn after tool call
|
|
236
|
+
let afterReflection: string | undefined;
|
|
237
|
+
let confidenceDelta: number | undefined;
|
|
238
|
+
|
|
239
|
+
if (tc.outcome === 'failure') {
|
|
240
|
+
const afterTurn = sortedTurns.find(
|
|
241
|
+
turn => parseTs(turn.createdAt) > tcTime
|
|
242
|
+
);
|
|
243
|
+
if (afterTurn) {
|
|
244
|
+
afterReflection = afterTurn.sanitizedText.slice(0, 300);
|
|
245
|
+
}
|
|
246
|
+
|
|
247
|
+
// Compute confidence delta if both before and after turns exist
|
|
248
|
+
if (beforeTurn && afterTurn) {
|
|
249
|
+
const beforeConfidence = confidenceToNumber(
|
|
250
|
+
mapConfidenceSignal(computeThinkingModelActivation(beforeTurn.sanitizedText))
|
|
251
|
+
);
|
|
252
|
+
const afterConfidence = confidenceToNumber(
|
|
253
|
+
mapConfidenceSignal(computeThinkingModelActivation(afterTurn.sanitizedText))
|
|
254
|
+
);
|
|
255
|
+
confidenceDelta = Math.round((afterConfidence - beforeConfidence) * 100) / 100;
|
|
256
|
+
}
|
|
257
|
+
}
|
|
258
|
+
|
|
259
|
+
const result: DerivedDecisionPoint = {
|
|
260
|
+
toolName: tc.toolName,
|
|
261
|
+
outcome: tc.outcome,
|
|
262
|
+
beforeContext,
|
|
263
|
+
};
|
|
264
|
+
if (afterReflection !== undefined) result.afterReflection = afterReflection;
|
|
265
|
+
if (confidenceDelta !== undefined) result.confidenceDelta = confidenceDelta;
|
|
266
|
+
return result;
|
|
267
|
+
});
|
|
268
|
+
}
|
|
269
|
+
|
|
270
|
+
// ---------------------------------------------------------------------------
|
|
271
|
+
// deriveContextualFactors (DERIV-03)
|
|
272
|
+
// ---------------------------------------------------------------------------
|
|
273
|
+
|
|
274
|
+
/**
|
|
275
|
+
* Compute contextual factors from session snapshot data.
|
|
276
|
+
*
|
|
277
|
+
* DERIV-03: Four boolean factors indicating the environment
|
|
278
|
+
* the agent was operating in. All derived from existing snapshot
|
|
279
|
+
* fields -- no schema changes.
|
|
280
|
+
*
|
|
281
|
+
* Empty/missing data returns all-false defaults. Never throws.
|
|
282
|
+
*/
|
|
283
|
+
export function deriveContextualFactors(
|
|
284
|
+
snapshot: NocturnalSessionSnapshot,
|
|
285
|
+
): DerivedContextualFactors {
|
|
286
|
+
const defaults: DerivedContextualFactors = {
|
|
287
|
+
fileStructureKnown: false,
|
|
288
|
+
errorHistoryPresent: false,
|
|
289
|
+
userGuidanceAvailable: false,
|
|
290
|
+
timePressure: false,
|
|
291
|
+
};
|
|
292
|
+
|
|
293
|
+
if (!snapshot) return defaults;
|
|
294
|
+
|
|
295
|
+
const { toolCalls = [], userTurns = [] } = snapshot;
|
|
296
|
+
|
|
297
|
+
// fileStructureKnown: any Read tool precedes any Write tool in chronological order
|
|
298
|
+
let fileStructureKnown = false;
|
|
299
|
+
const isReadTool = (name: string) => /^(read|grep|search|find|inspect|look)/i.test(name);
|
|
300
|
+
const isWriteTool = (name: string) => /^(edit|write|create|delete|remove|move|rename)/i.test(name);
|
|
301
|
+
let hasSeenRead = false;
|
|
302
|
+
for (const tc of toolCalls) {
|
|
303
|
+
if (isReadTool(tc.toolName)) hasSeenRead = true;
|
|
304
|
+
if (isWriteTool(tc.toolName) && hasSeenRead) {
|
|
305
|
+
fileStructureKnown = true;
|
|
306
|
+
break;
|
|
307
|
+
}
|
|
308
|
+
}
|
|
309
|
+
|
|
310
|
+
// errorHistoryPresent: any tool call with outcome === 'failure'
|
|
311
|
+
const errorHistoryPresent = toolCalls.some(tc => tc.outcome === 'failure');
|
|
312
|
+
|
|
313
|
+
// userGuidanceAvailable: any user turn with correctionDetected === true
|
|
314
|
+
const userGuidanceAvailable = (userTurns || []).some(ut => ut.correctionDetected === true);
|
|
315
|
+
|
|
316
|
+
// timePressure: >50% of consecutive tool call pairs have < 2s gap
|
|
317
|
+
let timePressure = false;
|
|
318
|
+
if (toolCalls.length >= 2) {
|
|
319
|
+
const sorted = [...toolCalls].sort(
|
|
320
|
+
(a, b) => parseTs(a.createdAt) - parseTs(b.createdAt)
|
|
321
|
+
);
|
|
322
|
+
let rapidGaps = 0;
|
|
323
|
+
for (let i = 0; i < sorted.length - 1; i++) {
|
|
324
|
+
const gap = parseTs(sorted[i + 1].createdAt) - parseTs(sorted[i].createdAt);
|
|
325
|
+
if (gap < 2000) rapidGaps++;
|
|
326
|
+
}
|
|
327
|
+
const totalPairs = sorted.length - 1;
|
|
328
|
+
timePressure = rapidGaps / totalPairs > 0.5;
|
|
329
|
+
}
|
|
330
|
+
|
|
331
|
+
return {
|
|
332
|
+
fileStructureKnown,
|
|
333
|
+
errorHistoryPresent,
|
|
334
|
+
userGuidanceAvailable,
|
|
335
|
+
timePressure,
|
|
336
|
+
};
|
|
337
|
+
}
|