@saulwade/swl-ses 1.3.7 → 1.4.0
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/CLAUDE.md +12 -4
- package/README.md +1 -1
- package/bin/swl-mcp-server.js +187 -187
- package/bin/swl-webhook-server.js +198 -0
- package/comandos/swl/.evolved.json +22 -22
- package/comandos/swl/adoptar-proyecto.md +21 -1
- package/comandos/swl/claudemd.md +14 -1
- package/comandos/swl/contribuir.md +233 -233
- package/comandos/swl/exportar-vault.md +207 -7
- package/comandos/swl/nuevo-proyecto.md +24 -2
- package/gateway/adapters/base.js +109 -0
- package/gateway/adapters/discord.js +167 -0
- package/gateway/adapters/email.js +221 -0
- package/gateway/adapters/slack.js +192 -0
- package/gateway/adapters/telegram.js +183 -0
- package/gateway/adapters/webhook.js +113 -0
- package/gateway/adapters/whatsapp.js +214 -0
- package/gateway/agent-executor.js +322 -0
- package/gateway/command-relay.js +271 -0
- package/gateway/cron/jobs.js +263 -0
- package/gateway/cron/scheduler.js +322 -0
- package/gateway/cron/store.js +335 -0
- package/gateway/index.js +320 -0
- package/gateway/lib/event-channel.js +191 -0
- package/gateway/session.js +131 -0
- package/gateway/webhook-server.js +324 -0
- package/habilidades/backend-production-resilience/SKILL.md +288 -288
- package/habilidades/benchmark-memoria/SKILL.md +186 -186
- package/habilidades/build-errors-nextjs/SKILL.md +55 -1
- package/habilidades/diagrama-arquitectura/assets/template.html +276 -276
- package/habilidades/doubt-driven-review/SKILL.md +171 -171
- package/habilidades/doubt-driven-review/recursos/EXAMPLES.md +130 -130
- package/habilidades/eval-framework/SKILL.md +212 -212
- package/habilidades/extractor-de-aprendizajes/SKILL.md +24 -10
- package/habilidades/harness-claude-code/SKILL.md +299 -299
- package/habilidades/infra-github-actions/SKILL.md +166 -166
- package/habilidades/legacy-code-rescue/SKILL.md +267 -267
- package/habilidades/manejo-errores/.evolved.json +8 -8
- package/habilidades/meta-skills-estandar/recursos/convencion-examples.md +93 -93
- package/habilidades/meta-skills-estandar/recursos/skills-as-agents.md +163 -163
- package/habilidades/nextjs-testing/SKILL.md +89 -5
- package/habilidades/node-experto/SKILL.md +37 -1
- package/habilidades/patrones-python/SKILL.md +229 -229
- package/habilidades/patrones-python/recursos/patrones-avanzados.md +469 -469
- package/habilidades/planear-fase/SKILL.md +319 -319
- package/habilidades/react-experto/SKILL.md +45 -4
- package/habilidades/release-semver/.evolved.json +8 -8
- package/habilidades/swl-claudemd/SKILL.md +15 -1
- package/habilidades/tdd-workflow/SKILL.md +36 -4
- package/habilidades/testing-python/SKILL.md +340 -340
- package/hooks/claudemd-bloat-detector.js +161 -161
- package/hooks/inyeccion-contexto.js +8 -3
- package/hooks/lib/agent-routing.js +107 -107
- package/hooks/lib/auto-consolidator.js +335 -335
- package/hooks/lib/error-classifier.js +308 -308
- package/hooks/lib/merkle-audit.js +96 -96
- package/hooks/lib/provenance-tracker.js +191 -191
- package/hooks/lib/rate-limit-ip.js +177 -0
- package/hooks/lib/rate-limit-tracker.js +253 -253
- package/hooks/lib/resource-quota.js +122 -122
- package/hooks/lib/retry-jitter.js +165 -165
- package/hooks/lib/skill-auditor.js +588 -588
- package/hooks/lib/sync-status.js +228 -228
- package/hooks/lib/taint-tracker.js +107 -107
- package/hooks/lib/text-similarity.js +241 -241
- package/hooks/lib/toon-compressor.js +245 -245
- package/hooks/lib/webhook-dedup.js +184 -0
- package/hooks/lib/webhook-verify.js +123 -0
- package/hooks/proteccion-rutas.js +120 -15
- package/hooks/registro-turnos.js +209 -209
- package/hooks/sugerir-regenerar-inventario.js +170 -170
- package/hooks/validar-formato-post-subagente.js +140 -140
- package/hooks/validar-memoria-hook.js +218 -218
- package/instintos/prompt-appendices.yaml +57 -57
- package/manifiestos/agent-output-schemas.json +57 -57
- package/manifiestos/modulos.json +1 -0
- package/manifiestos/skills-lock.json +37 -37
- package/package.json +5 -3
- package/plantillas/auditor-veto-template.md +105 -105
- package/plantillas/github-workflows/README.md +47 -47
- package/plantillas/github-workflows/release-please.yml +44 -44
- package/plantillas/github-workflows/swl-ci.yml +107 -107
- package/plantillas/github-workflows/swl-security.yml +51 -51
- package/plugin.json +1 -1
- package/reglas/analisis-previo-tareas-grandes.md +172 -172
- package/reglas/arreglar-al-detectar.md +147 -147
- package/reglas/fragmentos-compartidos.md +152 -152
- package/reglas/harness-claude-code.md +213 -213
- package/reglas/usar-context7.md +226 -226
- package/reglas/usar-sistema-swl.md +251 -0
- package/schemas/diary-entry.schema.json +80 -80
- package/scripts/benchmark-memoria.js +167 -167
- package/scripts/comandos/skills.js +251 -2
- package/scripts/configurar-branch-protection.js +418 -418
- package/scripts/detectar-aprendizajes-duplicados.js +151 -151
- package/scripts/field-report.js +199 -199
- package/scripts/generar-checklists-consolidados.js +273 -273
- package/scripts/generar-inventario.js +420 -420
- package/scripts/generar-matriz-lenguajes.js +271 -271
- package/scripts/lib/artefactos-python.js +43 -43
- package/scripts/lib/benchmark-metrics.js +160 -160
- package/scripts/lib/budget-enforcer.js +252 -252
- package/scripts/lib/configurar-ci.js +380 -380
- package/scripts/lib/contadores-inventario.js +217 -217
- package/scripts/lib/detectar-stack-detallado.js +307 -307
- package/scripts/lib/diary-entry.js +234 -234
- package/scripts/lib/eval-metrics-store.js +218 -218
- package/scripts/lib/eval-quality.js +171 -171
- package/scripts/lib/eval-schemas.js +144 -144
- package/scripts/lib/eval-self-correct.js +106 -106
- package/scripts/lib/eval-validator.js +185 -185
- package/scripts/lib/jaccard-similarity.js +98 -98
- package/scripts/lib/longmemeval-runner.js +125 -125
- package/scripts/lib/npm-version.js +261 -261
- package/scripts/lib/paquetes-conocidos.js +50 -50
- package/scripts/lib/prompt-builder.js +264 -264
- package/scripts/lib/rrf-fusion.js +175 -175
- package/scripts/lib/scoring-instintos.js +277 -277
- package/scripts/lib/semantic-search.js +252 -252
- package/scripts/limpiar-artefactos-python.js +131 -131
- package/scripts/mcp-server/README.md +128 -128
- package/scripts/mcp-server/handlers.js +206 -206
- package/scripts/migrar-csv-a-array.js +168 -168
- package/scripts/migrar-fase-dominio.js +201 -201
- package/scripts/publicar.js +511 -511
- package/scripts/run-eval.js +141 -141
- package/scripts/validar-manifest.js +195 -195
- package/scripts/validar-userland-vacio.js +110 -110
- package/scripts/verificar-release.js +110 -0
|
@@ -1,277 +1,277 @@
|
|
|
1
|
-
'use strict';
|
|
2
|
-
|
|
3
|
-
/**
|
|
4
|
-
* scoring-instintos.js
|
|
5
|
-
*
|
|
6
|
-
* Sistema de decay exponencial + maturity transitions para instintos.
|
|
7
|
-
*
|
|
8
|
-
* Patrón adaptado de ACE (Agentic Context Engineering) en
|
|
9
|
-
* `temp/cass_memory_system-main/src/scoring.ts`. Diferencias:
|
|
10
|
-
* - Funciones puras zero-deps (Node stdlib only).
|
|
11
|
-
* - Backward compatible: instintos sin los campos nuevos siguen siendo válidos.
|
|
12
|
-
* - No persiste — el caller decide cuándo escribir el resultado.
|
|
13
|
-
*
|
|
14
|
-
* Modelo:
|
|
15
|
-
* - `confidence` original es estática (asignada al crear).
|
|
16
|
-
* - `effective_confidence` se calcula dinámicamente:
|
|
17
|
-
* confidence × decay_factor − harmful_penalty
|
|
18
|
-
* - `decay_factor` = 2^(-days_since_validated / half_life_days) (default 90d).
|
|
19
|
-
* - `harmful_penalty` = HARMFUL_PENALTY_WEIGHT × harmful_ratio.
|
|
20
|
-
* - `maturity` se deriva de effective_confidence + evidence_count + harmful_ratio.
|
|
21
|
-
* - `shouldAutoDeprecate` cuando harmful_ratio > HARMFUL_RATIO_THRESHOLD
|
|
22
|
-
* con al menos MIN_FEEDBACK_FOR_DEPRECATION eventos.
|
|
23
|
-
*
|
|
24
|
-
* Campos NUEVOS opcionales en instintos (todos con defaults seguros):
|
|
25
|
-
* - decay_half_life_days (default: 90)
|
|
26
|
-
* - helpful_count (default: 0)
|
|
27
|
-
* - harmful_count (default: 0)
|
|
28
|
-
* - last_validated_at (default: last_validated, o created_at)
|
|
29
|
-
*
|
|
30
|
-
* @module scripts/lib/scoring-instintos
|
|
31
|
-
*/
|
|
32
|
-
|
|
33
|
-
// ── constantes ────────────────────────────────────────────────────────────────
|
|
34
|
-
|
|
35
|
-
const HALF_LIFE_DEFAULT_DAYS = 90;
|
|
36
|
-
const HARMFUL_RATIO_THRESHOLD = 0.3;
|
|
37
|
-
const MIN_FEEDBACK_FOR_DEPRECATION = 3;
|
|
38
|
-
const HARMFUL_PENALTY_WEIGHT = 0.5;
|
|
39
|
-
const PROVEN_CONFIDENCE_THRESHOLD = 0.7;
|
|
40
|
-
const ESTABLISHED_THRESHOLD = 0.5;
|
|
41
|
-
const PROVEN_EVIDENCE_THRESHOLD = 3;
|
|
42
|
-
// Modelo de reforzamiento por feedback positivo (estilo agentmemory/lessons.ts).
|
|
43
|
-
// Cada feedback positivo cierra un (1 - REINFORCEMENT_DECAY) = 10% del gap
|
|
44
|
-
// entre la confianza actual y 1.0 — diminishing returns naturales.
|
|
45
|
-
const REINFORCEMENT_DECAY = 0.9;
|
|
46
|
-
|
|
47
|
-
// ── helpers ───────────────────────────────────────────────────────────────────
|
|
48
|
-
|
|
49
|
-
/**
|
|
50
|
-
* Días enteros entre dos fechas YYYY-MM-DD (o Date).
|
|
51
|
-
* Tolerante a formato inválido: devuelve 0 si no parsea.
|
|
52
|
-
*/
|
|
53
|
-
function daysBetween(d1, d2) {
|
|
54
|
-
const t1 = d1 instanceof Date ? d1.getTime() : Date.parse(String(d1));
|
|
55
|
-
const t2 = d2 instanceof Date ? d2.getTime() : Date.parse(String(d2));
|
|
56
|
-
if (Number.isNaN(t1) || Number.isNaN(t2)) return 0;
|
|
57
|
-
const diffMs = Math.abs(t2 - t1);
|
|
58
|
-
return Math.floor(diffMs / (24 * 3600 * 1000));
|
|
59
|
-
}
|
|
60
|
-
|
|
61
|
-
function clamp(n, min, max) {
|
|
62
|
-
if (Number.isNaN(n)) return min;
|
|
63
|
-
return Math.max(min, Math.min(max, n));
|
|
64
|
-
}
|
|
65
|
-
|
|
66
|
-
function getValidatedDate(instinto) {
|
|
67
|
-
return instinto.last_validated_at
|
|
68
|
-
|| instinto.last_validated
|
|
69
|
-
|| instinto.created_at
|
|
70
|
-
|| null;
|
|
71
|
-
}
|
|
72
|
-
|
|
73
|
-
// ── funciones puras de scoring ────────────────────────────────────────────────
|
|
74
|
-
|
|
75
|
-
/**
|
|
76
|
-
* Factor de decay exponencial.
|
|
77
|
-
*
|
|
78
|
-
* decay(0) = 1.0
|
|
79
|
-
* decay(half) = 0.5
|
|
80
|
-
* decay(2×half) = 0.25
|
|
81
|
-
*
|
|
82
|
-
* @param {number} daysSinceValidated
|
|
83
|
-
* @param {number} [halfLifeDays=90]
|
|
84
|
-
* @returns {number} factor en [0, 1]
|
|
85
|
-
*/
|
|
86
|
-
function decayFactor(daysSinceValidated, halfLifeDays = HALF_LIFE_DEFAULT_DAYS) {
|
|
87
|
-
if (daysSinceValidated <= 0) return 1.0;
|
|
88
|
-
if (halfLifeDays <= 0) return 1.0;
|
|
89
|
-
return Math.pow(2, -daysSinceValidated / halfLifeDays);
|
|
90
|
-
}
|
|
91
|
-
|
|
92
|
-
/**
|
|
93
|
-
* Ratio de feedback dañino sobre el total de feedback recibido.
|
|
94
|
-
* Devuelve 0 si no hay feedback.
|
|
95
|
-
*/
|
|
96
|
-
function harmfulRatio(instinto) {
|
|
97
|
-
const helpful = instinto.helpful_count || 0;
|
|
98
|
-
const harmful = instinto.harmful_count || 0;
|
|
99
|
-
const total = helpful + harmful;
|
|
100
|
-
if (total === 0) return 0;
|
|
101
|
-
return harmful / total;
|
|
102
|
-
}
|
|
103
|
-
|
|
104
|
-
/**
|
|
105
|
-
* Confianza reforzada por feedback positivo acumulado.
|
|
106
|
-
*
|
|
107
|
-
* Aplica un modelo de diminishing returns inspirado en `lessons.ts` de
|
|
108
|
-
* agentmemory: cada feedback positivo cierra una fracción del gap entre la
|
|
109
|
-
* confianza actual y 1.0. Aplicado N veces sobre la confianza base:
|
|
110
|
-
*
|
|
111
|
-
* c_N = 1 - REINFORCEMENT_DECAY^N · (1 - c_0)
|
|
112
|
-
*
|
|
113
|
-
* Con `REINFORCEMENT_DECAY = 0.9`:
|
|
114
|
-
* - 1 feedback: c_0=0.5 → 0.55
|
|
115
|
-
* - 5 feedback: c_0=0.5 → 0.705
|
|
116
|
-
* - 10 feedback: c_0=0.5 → 0.826
|
|
117
|
-
* - 50 feedback: c_0=0.5 → 0.997
|
|
118
|
-
*
|
|
119
|
-
* Si no hay `helpful_count` (o es 0), devuelve la confianza base sin tocar.
|
|
120
|
-
* Backward compatible: instintos sin el campo siguen comportándose igual.
|
|
121
|
-
*
|
|
122
|
-
* No persiste el resultado — se computa on-demand desde `helpful_count`. Eso
|
|
123
|
-
* preserva el invariante "confidence original es estática".
|
|
124
|
-
*/
|
|
125
|
-
function reinforcedConfidence(instinto) {
|
|
126
|
-
const baseConfidence = clamp(instinto.confidence || 0, 0, 1);
|
|
127
|
-
const helpful = instinto.helpful_count || 0;
|
|
128
|
-
if (helpful <= 0) return baseConfidence;
|
|
129
|
-
return 1 - Math.pow(REINFORCEMENT_DECAY, helpful) * (1 - baseConfidence);
|
|
130
|
-
}
|
|
131
|
-
|
|
132
|
-
/**
|
|
133
|
-
* Confianza efectiva considerando reforzamiento, decay temporal y feedback
|
|
134
|
-
* negativo.
|
|
135
|
-
*
|
|
136
|
-
* effective = reinforced × decay − harmful_penalty
|
|
137
|
-
*
|
|
138
|
-
* @param {object} instinto
|
|
139
|
-
* @param {string|Date} [currentDate=now] — fecha de referencia
|
|
140
|
-
* @returns {number} en [0, 1]
|
|
141
|
-
*/
|
|
142
|
-
function effectiveConfidence(instinto, currentDate) {
|
|
143
|
-
const now = currentDate || new Date();
|
|
144
|
-
const validatedAt = getValidatedDate(instinto);
|
|
145
|
-
const halfLife = instinto.decay_half_life_days || HALF_LIFE_DEFAULT_DAYS;
|
|
146
|
-
|
|
147
|
-
const days = validatedAt ? daysBetween(validatedAt, now) : 0;
|
|
148
|
-
const decay = decayFactor(days, halfLife);
|
|
149
|
-
const reinforced = reinforcedConfidence(instinto);
|
|
150
|
-
const penalty = HARMFUL_PENALTY_WEIGHT * harmfulRatio(instinto);
|
|
151
|
-
|
|
152
|
-
return clamp(reinforced * decay - penalty, 0, 1);
|
|
153
|
-
}
|
|
154
|
-
|
|
155
|
-
/**
|
|
156
|
-
* Si el feedback negativo supera el umbral, recomendar auto-deprecación.
|
|
157
|
-
* Requiere al menos MIN_FEEDBACK_FOR_DEPRECATION eventos para evitar
|
|
158
|
-
* sobre-reaccionar a un solo voto.
|
|
159
|
-
*/
|
|
160
|
-
function shouldAutoDeprecate(instinto) {
|
|
161
|
-
const helpful = instinto.helpful_count || 0;
|
|
162
|
-
const harmful = instinto.harmful_count || 0;
|
|
163
|
-
const total = helpful + harmful;
|
|
164
|
-
if (total < MIN_FEEDBACK_FOR_DEPRECATION) return false;
|
|
165
|
-
return harmfulRatio(instinto) > HARMFUL_RATIO_THRESHOLD;
|
|
166
|
-
}
|
|
167
|
-
|
|
168
|
-
/**
|
|
169
|
-
* Estado de maduración derivado de effective_confidence + evidencia + feedback.
|
|
170
|
-
*
|
|
171
|
-
* Estados:
|
|
172
|
-
* - candidate : effective < 0.5 (recién aprendido o decaído)
|
|
173
|
-
* - established : 0.5 ≤ effective < 0.7
|
|
174
|
-
* - proven : effective ≥ 0.7 AND evidence_count ≥ 3
|
|
175
|
-
* - deprecated : shouldAutoDeprecate, status === 'archived', o
|
|
176
|
-
* status === 'degraded' con contradictions ≥ 3
|
|
177
|
-
*/
|
|
178
|
-
function maturityState(instinto, currentDate) {
|
|
179
|
-
if (instinto.status === 'archived') return 'deprecated';
|
|
180
|
-
if (shouldAutoDeprecate(instinto)) return 'deprecated';
|
|
181
|
-
if (instinto.status === 'degraded' && (instinto.contradictions || 0) >= 3) {
|
|
182
|
-
return 'deprecated';
|
|
183
|
-
}
|
|
184
|
-
|
|
185
|
-
const eff = effectiveConfidence(instinto, currentDate);
|
|
186
|
-
const evidence = instinto.evidence_count || 0;
|
|
187
|
-
|
|
188
|
-
if (eff >= PROVEN_CONFIDENCE_THRESHOLD && evidence >= PROVEN_EVIDENCE_THRESHOLD) {
|
|
189
|
-
return 'proven';
|
|
190
|
-
}
|
|
191
|
-
if (eff >= ESTABLISHED_THRESHOLD) return 'established';
|
|
192
|
-
return 'candidate';
|
|
193
|
-
}
|
|
194
|
-
|
|
195
|
-
/**
|
|
196
|
-
* Aplicar feedback al instinto (pure: devuelve copia).
|
|
197
|
-
* type: 'helpful' | 'harmful'
|
|
198
|
-
*
|
|
199
|
-
* Refresca last_validated_at — un feedback es una validación implícita.
|
|
200
|
-
*/
|
|
201
|
-
function applyFeedback(instinto, type, currentDate) {
|
|
202
|
-
if (type !== 'helpful' && type !== 'harmful') {
|
|
203
|
-
throw new Error(`Tipo de feedback inválido: ${type}`);
|
|
204
|
-
}
|
|
205
|
-
const now = (currentDate instanceof Date)
|
|
206
|
-
? currentDate.toISOString().slice(0, 10)
|
|
207
|
-
: (currentDate || new Date().toISOString().slice(0, 10));
|
|
208
|
-
|
|
209
|
-
const next = { ...instinto };
|
|
210
|
-
if (type === 'helpful') next.helpful_count = (next.helpful_count || 0) + 1;
|
|
211
|
-
else next.harmful_count = (next.harmful_count || 0) + 1;
|
|
212
|
-
next.last_validated_at = now;
|
|
213
|
-
return next;
|
|
214
|
-
}
|
|
215
|
-
|
|
216
|
-
/**
|
|
217
|
-
* Recomputa status legacy + maturity como vista unificada.
|
|
218
|
-
* Pure: devuelve un objeto enriquecido sin mutar el original.
|
|
219
|
-
*
|
|
220
|
-
* Si shouldAutoDeprecate, propone status='degraded' (no archive — el archive
|
|
221
|
-
* es decisión humana). El caller decide si persistir el cambio.
|
|
222
|
-
*/
|
|
223
|
-
function recomputeStatus(instinto, currentDate) {
|
|
224
|
-
const maturity = maturityState(instinto, currentDate);
|
|
225
|
-
const effective = effectiveConfidence(instinto, currentDate);
|
|
226
|
-
const proposed = { ...instinto, maturity, effective_confidence: Math.round(effective * 1000) / 1000 };
|
|
227
|
-
|
|
228
|
-
if (shouldAutoDeprecate(instinto) && instinto.status !== 'archived') {
|
|
229
|
-
proposed.status_proposed = 'degraded';
|
|
230
|
-
proposed.deprecation_reason = `harmful_ratio=${harmfulRatio(instinto).toFixed(2)} con ${(instinto.helpful_count||0)+(instinto.harmful_count||0)} eventos`;
|
|
231
|
-
}
|
|
232
|
-
|
|
233
|
-
return proposed;
|
|
234
|
-
}
|
|
235
|
-
|
|
236
|
-
/**
|
|
237
|
-
* Adjuntar source tracing a un instinto (pure: devuelve copia).
|
|
238
|
-
* Mantiene unicidad y orden de inserción.
|
|
239
|
-
*/
|
|
240
|
-
function addSource(instinto, { sessionId, agentName } = {}) {
|
|
241
|
-
const next = { ...instinto };
|
|
242
|
-
if (sessionId) {
|
|
243
|
-
const s = new Set(next.source_sessions || []);
|
|
244
|
-
s.add(sessionId);
|
|
245
|
-
next.source_sessions = [...s];
|
|
246
|
-
}
|
|
247
|
-
if (agentName) {
|
|
248
|
-
const a = new Set(next.source_agents || []);
|
|
249
|
-
a.add(agentName);
|
|
250
|
-
next.source_agents = [...a];
|
|
251
|
-
}
|
|
252
|
-
return next;
|
|
253
|
-
}
|
|
254
|
-
|
|
255
|
-
// ── exports ───────────────────────────────────────────────────────────────────
|
|
256
|
-
|
|
257
|
-
module.exports = {
|
|
258
|
-
daysBetween,
|
|
259
|
-
decayFactor,
|
|
260
|
-
harmfulRatio,
|
|
261
|
-
reinforcedConfidence,
|
|
262
|
-
effectiveConfidence,
|
|
263
|
-
shouldAutoDeprecate,
|
|
264
|
-
maturityState,
|
|
265
|
-
applyFeedback,
|
|
266
|
-
recomputeStatus,
|
|
267
|
-
addSource,
|
|
268
|
-
// constantes exportadas para tests y configuración
|
|
269
|
-
HALF_LIFE_DEFAULT_DAYS,
|
|
270
|
-
HARMFUL_RATIO_THRESHOLD,
|
|
271
|
-
MIN_FEEDBACK_FOR_DEPRECATION,
|
|
272
|
-
HARMFUL_PENALTY_WEIGHT,
|
|
273
|
-
PROVEN_CONFIDENCE_THRESHOLD,
|
|
274
|
-
ESTABLISHED_THRESHOLD,
|
|
275
|
-
PROVEN_EVIDENCE_THRESHOLD,
|
|
276
|
-
REINFORCEMENT_DECAY,
|
|
277
|
-
};
|
|
1
|
+
'use strict';
|
|
2
|
+
|
|
3
|
+
/**
|
|
4
|
+
* scoring-instintos.js
|
|
5
|
+
*
|
|
6
|
+
* Sistema de decay exponencial + maturity transitions para instintos.
|
|
7
|
+
*
|
|
8
|
+
* Patrón adaptado de ACE (Agentic Context Engineering) en
|
|
9
|
+
* `temp/cass_memory_system-main/src/scoring.ts`. Diferencias:
|
|
10
|
+
* - Funciones puras zero-deps (Node stdlib only).
|
|
11
|
+
* - Backward compatible: instintos sin los campos nuevos siguen siendo válidos.
|
|
12
|
+
* - No persiste — el caller decide cuándo escribir el resultado.
|
|
13
|
+
*
|
|
14
|
+
* Modelo:
|
|
15
|
+
* - `confidence` original es estática (asignada al crear).
|
|
16
|
+
* - `effective_confidence` se calcula dinámicamente:
|
|
17
|
+
* confidence × decay_factor − harmful_penalty
|
|
18
|
+
* - `decay_factor` = 2^(-days_since_validated / half_life_days) (default 90d).
|
|
19
|
+
* - `harmful_penalty` = HARMFUL_PENALTY_WEIGHT × harmful_ratio.
|
|
20
|
+
* - `maturity` se deriva de effective_confidence + evidence_count + harmful_ratio.
|
|
21
|
+
* - `shouldAutoDeprecate` cuando harmful_ratio > HARMFUL_RATIO_THRESHOLD
|
|
22
|
+
* con al menos MIN_FEEDBACK_FOR_DEPRECATION eventos.
|
|
23
|
+
*
|
|
24
|
+
* Campos NUEVOS opcionales en instintos (todos con defaults seguros):
|
|
25
|
+
* - decay_half_life_days (default: 90)
|
|
26
|
+
* - helpful_count (default: 0)
|
|
27
|
+
* - harmful_count (default: 0)
|
|
28
|
+
* - last_validated_at (default: last_validated, o created_at)
|
|
29
|
+
*
|
|
30
|
+
* @module scripts/lib/scoring-instintos
|
|
31
|
+
*/
|
|
32
|
+
|
|
33
|
+
// ── constantes ────────────────────────────────────────────────────────────────
|
|
34
|
+
|
|
35
|
+
const HALF_LIFE_DEFAULT_DAYS = 90;
|
|
36
|
+
const HARMFUL_RATIO_THRESHOLD = 0.3;
|
|
37
|
+
const MIN_FEEDBACK_FOR_DEPRECATION = 3;
|
|
38
|
+
const HARMFUL_PENALTY_WEIGHT = 0.5;
|
|
39
|
+
const PROVEN_CONFIDENCE_THRESHOLD = 0.7;
|
|
40
|
+
const ESTABLISHED_THRESHOLD = 0.5;
|
|
41
|
+
const PROVEN_EVIDENCE_THRESHOLD = 3;
|
|
42
|
+
// Modelo de reforzamiento por feedback positivo (estilo agentmemory/lessons.ts).
|
|
43
|
+
// Cada feedback positivo cierra un (1 - REINFORCEMENT_DECAY) = 10% del gap
|
|
44
|
+
// entre la confianza actual y 1.0 — diminishing returns naturales.
|
|
45
|
+
const REINFORCEMENT_DECAY = 0.9;
|
|
46
|
+
|
|
47
|
+
// ── helpers ───────────────────────────────────────────────────────────────────
|
|
48
|
+
|
|
49
|
+
/**
|
|
50
|
+
* Días enteros entre dos fechas YYYY-MM-DD (o Date).
|
|
51
|
+
* Tolerante a formato inválido: devuelve 0 si no parsea.
|
|
52
|
+
*/
|
|
53
|
+
function daysBetween(d1, d2) {
|
|
54
|
+
const t1 = d1 instanceof Date ? d1.getTime() : Date.parse(String(d1));
|
|
55
|
+
const t2 = d2 instanceof Date ? d2.getTime() : Date.parse(String(d2));
|
|
56
|
+
if (Number.isNaN(t1) || Number.isNaN(t2)) return 0;
|
|
57
|
+
const diffMs = Math.abs(t2 - t1);
|
|
58
|
+
return Math.floor(diffMs / (24 * 3600 * 1000));
|
|
59
|
+
}
|
|
60
|
+
|
|
61
|
+
function clamp(n, min, max) {
|
|
62
|
+
if (Number.isNaN(n)) return min;
|
|
63
|
+
return Math.max(min, Math.min(max, n));
|
|
64
|
+
}
|
|
65
|
+
|
|
66
|
+
function getValidatedDate(instinto) {
|
|
67
|
+
return instinto.last_validated_at
|
|
68
|
+
|| instinto.last_validated
|
|
69
|
+
|| instinto.created_at
|
|
70
|
+
|| null;
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
// ── funciones puras de scoring ────────────────────────────────────────────────
|
|
74
|
+
|
|
75
|
+
/**
|
|
76
|
+
* Factor de decay exponencial.
|
|
77
|
+
*
|
|
78
|
+
* decay(0) = 1.0
|
|
79
|
+
* decay(half) = 0.5
|
|
80
|
+
* decay(2×half) = 0.25
|
|
81
|
+
*
|
|
82
|
+
* @param {number} daysSinceValidated
|
|
83
|
+
* @param {number} [halfLifeDays=90]
|
|
84
|
+
* @returns {number} factor en [0, 1]
|
|
85
|
+
*/
|
|
86
|
+
function decayFactor(daysSinceValidated, halfLifeDays = HALF_LIFE_DEFAULT_DAYS) {
|
|
87
|
+
if (daysSinceValidated <= 0) return 1.0;
|
|
88
|
+
if (halfLifeDays <= 0) return 1.0;
|
|
89
|
+
return Math.pow(2, -daysSinceValidated / halfLifeDays);
|
|
90
|
+
}
|
|
91
|
+
|
|
92
|
+
/**
|
|
93
|
+
* Ratio de feedback dañino sobre el total de feedback recibido.
|
|
94
|
+
* Devuelve 0 si no hay feedback.
|
|
95
|
+
*/
|
|
96
|
+
function harmfulRatio(instinto) {
|
|
97
|
+
const helpful = instinto.helpful_count || 0;
|
|
98
|
+
const harmful = instinto.harmful_count || 0;
|
|
99
|
+
const total = helpful + harmful;
|
|
100
|
+
if (total === 0) return 0;
|
|
101
|
+
return harmful / total;
|
|
102
|
+
}
|
|
103
|
+
|
|
104
|
+
/**
|
|
105
|
+
* Confianza reforzada por feedback positivo acumulado.
|
|
106
|
+
*
|
|
107
|
+
* Aplica un modelo de diminishing returns inspirado en `lessons.ts` de
|
|
108
|
+
* agentmemory: cada feedback positivo cierra una fracción del gap entre la
|
|
109
|
+
* confianza actual y 1.0. Aplicado N veces sobre la confianza base:
|
|
110
|
+
*
|
|
111
|
+
* c_N = 1 - REINFORCEMENT_DECAY^N · (1 - c_0)
|
|
112
|
+
*
|
|
113
|
+
* Con `REINFORCEMENT_DECAY = 0.9`:
|
|
114
|
+
* - 1 feedback: c_0=0.5 → 0.55
|
|
115
|
+
* - 5 feedback: c_0=0.5 → 0.705
|
|
116
|
+
* - 10 feedback: c_0=0.5 → 0.826
|
|
117
|
+
* - 50 feedback: c_0=0.5 → 0.997
|
|
118
|
+
*
|
|
119
|
+
* Si no hay `helpful_count` (o es 0), devuelve la confianza base sin tocar.
|
|
120
|
+
* Backward compatible: instintos sin el campo siguen comportándose igual.
|
|
121
|
+
*
|
|
122
|
+
* No persiste el resultado — se computa on-demand desde `helpful_count`. Eso
|
|
123
|
+
* preserva el invariante "confidence original es estática".
|
|
124
|
+
*/
|
|
125
|
+
function reinforcedConfidence(instinto) {
|
|
126
|
+
const baseConfidence = clamp(instinto.confidence || 0, 0, 1);
|
|
127
|
+
const helpful = instinto.helpful_count || 0;
|
|
128
|
+
if (helpful <= 0) return baseConfidence;
|
|
129
|
+
return 1 - Math.pow(REINFORCEMENT_DECAY, helpful) * (1 - baseConfidence);
|
|
130
|
+
}
|
|
131
|
+
|
|
132
|
+
/**
|
|
133
|
+
* Confianza efectiva considerando reforzamiento, decay temporal y feedback
|
|
134
|
+
* negativo.
|
|
135
|
+
*
|
|
136
|
+
* effective = reinforced × decay − harmful_penalty
|
|
137
|
+
*
|
|
138
|
+
* @param {object} instinto
|
|
139
|
+
* @param {string|Date} [currentDate=now] — fecha de referencia
|
|
140
|
+
* @returns {number} en [0, 1]
|
|
141
|
+
*/
|
|
142
|
+
function effectiveConfidence(instinto, currentDate) {
|
|
143
|
+
const now = currentDate || new Date();
|
|
144
|
+
const validatedAt = getValidatedDate(instinto);
|
|
145
|
+
const halfLife = instinto.decay_half_life_days || HALF_LIFE_DEFAULT_DAYS;
|
|
146
|
+
|
|
147
|
+
const days = validatedAt ? daysBetween(validatedAt, now) : 0;
|
|
148
|
+
const decay = decayFactor(days, halfLife);
|
|
149
|
+
const reinforced = reinforcedConfidence(instinto);
|
|
150
|
+
const penalty = HARMFUL_PENALTY_WEIGHT * harmfulRatio(instinto);
|
|
151
|
+
|
|
152
|
+
return clamp(reinforced * decay - penalty, 0, 1);
|
|
153
|
+
}
|
|
154
|
+
|
|
155
|
+
/**
|
|
156
|
+
* Si el feedback negativo supera el umbral, recomendar auto-deprecación.
|
|
157
|
+
* Requiere al menos MIN_FEEDBACK_FOR_DEPRECATION eventos para evitar
|
|
158
|
+
* sobre-reaccionar a un solo voto.
|
|
159
|
+
*/
|
|
160
|
+
function shouldAutoDeprecate(instinto) {
|
|
161
|
+
const helpful = instinto.helpful_count || 0;
|
|
162
|
+
const harmful = instinto.harmful_count || 0;
|
|
163
|
+
const total = helpful + harmful;
|
|
164
|
+
if (total < MIN_FEEDBACK_FOR_DEPRECATION) return false;
|
|
165
|
+
return harmfulRatio(instinto) > HARMFUL_RATIO_THRESHOLD;
|
|
166
|
+
}
|
|
167
|
+
|
|
168
|
+
/**
|
|
169
|
+
* Estado de maduración derivado de effective_confidence + evidencia + feedback.
|
|
170
|
+
*
|
|
171
|
+
* Estados:
|
|
172
|
+
* - candidate : effective < 0.5 (recién aprendido o decaído)
|
|
173
|
+
* - established : 0.5 ≤ effective < 0.7
|
|
174
|
+
* - proven : effective ≥ 0.7 AND evidence_count ≥ 3
|
|
175
|
+
* - deprecated : shouldAutoDeprecate, status === 'archived', o
|
|
176
|
+
* status === 'degraded' con contradictions ≥ 3
|
|
177
|
+
*/
|
|
178
|
+
function maturityState(instinto, currentDate) {
|
|
179
|
+
if (instinto.status === 'archived') return 'deprecated';
|
|
180
|
+
if (shouldAutoDeprecate(instinto)) return 'deprecated';
|
|
181
|
+
if (instinto.status === 'degraded' && (instinto.contradictions || 0) >= 3) {
|
|
182
|
+
return 'deprecated';
|
|
183
|
+
}
|
|
184
|
+
|
|
185
|
+
const eff = effectiveConfidence(instinto, currentDate);
|
|
186
|
+
const evidence = instinto.evidence_count || 0;
|
|
187
|
+
|
|
188
|
+
if (eff >= PROVEN_CONFIDENCE_THRESHOLD && evidence >= PROVEN_EVIDENCE_THRESHOLD) {
|
|
189
|
+
return 'proven';
|
|
190
|
+
}
|
|
191
|
+
if (eff >= ESTABLISHED_THRESHOLD) return 'established';
|
|
192
|
+
return 'candidate';
|
|
193
|
+
}
|
|
194
|
+
|
|
195
|
+
/**
|
|
196
|
+
* Aplicar feedback al instinto (pure: devuelve copia).
|
|
197
|
+
* type: 'helpful' | 'harmful'
|
|
198
|
+
*
|
|
199
|
+
* Refresca last_validated_at — un feedback es una validación implícita.
|
|
200
|
+
*/
|
|
201
|
+
function applyFeedback(instinto, type, currentDate) {
|
|
202
|
+
if (type !== 'helpful' && type !== 'harmful') {
|
|
203
|
+
throw new Error(`Tipo de feedback inválido: ${type}`);
|
|
204
|
+
}
|
|
205
|
+
const now = (currentDate instanceof Date)
|
|
206
|
+
? currentDate.toISOString().slice(0, 10)
|
|
207
|
+
: (currentDate || new Date().toISOString().slice(0, 10));
|
|
208
|
+
|
|
209
|
+
const next = { ...instinto };
|
|
210
|
+
if (type === 'helpful') next.helpful_count = (next.helpful_count || 0) + 1;
|
|
211
|
+
else next.harmful_count = (next.harmful_count || 0) + 1;
|
|
212
|
+
next.last_validated_at = now;
|
|
213
|
+
return next;
|
|
214
|
+
}
|
|
215
|
+
|
|
216
|
+
/**
|
|
217
|
+
* Recomputa status legacy + maturity como vista unificada.
|
|
218
|
+
* Pure: devuelve un objeto enriquecido sin mutar el original.
|
|
219
|
+
*
|
|
220
|
+
* Si shouldAutoDeprecate, propone status='degraded' (no archive — el archive
|
|
221
|
+
* es decisión humana). El caller decide si persistir el cambio.
|
|
222
|
+
*/
|
|
223
|
+
function recomputeStatus(instinto, currentDate) {
|
|
224
|
+
const maturity = maturityState(instinto, currentDate);
|
|
225
|
+
const effective = effectiveConfidence(instinto, currentDate);
|
|
226
|
+
const proposed = { ...instinto, maturity, effective_confidence: Math.round(effective * 1000) / 1000 };
|
|
227
|
+
|
|
228
|
+
if (shouldAutoDeprecate(instinto) && instinto.status !== 'archived') {
|
|
229
|
+
proposed.status_proposed = 'degraded';
|
|
230
|
+
proposed.deprecation_reason = `harmful_ratio=${harmfulRatio(instinto).toFixed(2)} con ${(instinto.helpful_count||0)+(instinto.harmful_count||0)} eventos`;
|
|
231
|
+
}
|
|
232
|
+
|
|
233
|
+
return proposed;
|
|
234
|
+
}
|
|
235
|
+
|
|
236
|
+
/**
|
|
237
|
+
* Adjuntar source tracing a un instinto (pure: devuelve copia).
|
|
238
|
+
* Mantiene unicidad y orden de inserción.
|
|
239
|
+
*/
|
|
240
|
+
function addSource(instinto, { sessionId, agentName } = {}) {
|
|
241
|
+
const next = { ...instinto };
|
|
242
|
+
if (sessionId) {
|
|
243
|
+
const s = new Set(next.source_sessions || []);
|
|
244
|
+
s.add(sessionId);
|
|
245
|
+
next.source_sessions = [...s];
|
|
246
|
+
}
|
|
247
|
+
if (agentName) {
|
|
248
|
+
const a = new Set(next.source_agents || []);
|
|
249
|
+
a.add(agentName);
|
|
250
|
+
next.source_agents = [...a];
|
|
251
|
+
}
|
|
252
|
+
return next;
|
|
253
|
+
}
|
|
254
|
+
|
|
255
|
+
// ── exports ───────────────────────────────────────────────────────────────────
|
|
256
|
+
|
|
257
|
+
module.exports = {
|
|
258
|
+
daysBetween,
|
|
259
|
+
decayFactor,
|
|
260
|
+
harmfulRatio,
|
|
261
|
+
reinforcedConfidence,
|
|
262
|
+
effectiveConfidence,
|
|
263
|
+
shouldAutoDeprecate,
|
|
264
|
+
maturityState,
|
|
265
|
+
applyFeedback,
|
|
266
|
+
recomputeStatus,
|
|
267
|
+
addSource,
|
|
268
|
+
// constantes exportadas para tests y configuración
|
|
269
|
+
HALF_LIFE_DEFAULT_DAYS,
|
|
270
|
+
HARMFUL_RATIO_THRESHOLD,
|
|
271
|
+
MIN_FEEDBACK_FOR_DEPRECATION,
|
|
272
|
+
HARMFUL_PENALTY_WEIGHT,
|
|
273
|
+
PROVEN_CONFIDENCE_THRESHOLD,
|
|
274
|
+
ESTABLISHED_THRESHOLD,
|
|
275
|
+
PROVEN_EVIDENCE_THRESHOLD,
|
|
276
|
+
REINFORCEMENT_DECAY,
|
|
277
|
+
};
|