feat(observabilidade): tabela patterns + 6 detectores SQL

Adiciona tabela 'patterns' à BD sessions (UNIQUE por week_iso+pattern_key)
e helpers upsertPattern/getPatternsByWeek/getConsecutiveWeeks no SessionsDb.

Módulo patterns.ts implementa 6 detectores heurísticos para deteccão semanal:
  1. skills_with_high_error_rate (ratio > 0.2, severity warning|action)
  2. tools_low_efficiency (tool_calls/event_count médio > 0.5)
  3. skill_tool_pairs (top 5 co-ocorrências)
  4. duration_outliers (sessões > p95 com outcome != completed)
  5. abandoned_sessions (event_count<3 AND outcome=unknown, >=5)
  6. growing_complexity (avg tool_calls actual > anterior*1.3)

5 testes cobrem detector de erro, abandonadas, consecutive_weeks,
idempotência do upsert e toPatternRecord.

Refs Fase 6A · Desk #2059 · Project #65

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-04-23 02:17:21 +01:00
parent 2c8525bc8a
commit 2a523a505e
3 changed files with 568 additions and 0 deletions
+326
View File
@@ -0,0 +1,326 @@
/**
* Detector automático de padrões sobre a BD `sessions` (Observabilidade Fase 6A).
*
* Seis detectores heurísticos em SQL puro (via better-sqlite3). Cada detector
* devolve zero ou mais `Pattern` para a semana analisada. Pipeline:
* 1. Correr detectores sobre intervalo [weekStart, weekEnd]
* 2. Persistir via `upsertPattern` (idempotente por (week_iso, pattern_key))
* 3. Calcular `consecutive_weeks` olhando para semanas anteriores
*/
import type Database from 'better-sqlite3'
import type { SessionsDb, PatternRecord } from './db.js'
export type Severity = 'info' | 'warning' | 'action'
export interface Pattern {
pattern_key: string
title: string
description: string
severity: Severity
metric_value: number | null
sample_session_ids: string[]
affected_count: number
}
export interface DetectCtx {
db: Database.Database
weekStartIso: string
weekEndIso: string
}
/** Converte Date para string ISO UTC. */
function iso(d: Date): string {
return d.toISOString()
}
/**
* Calcula intervalo [segunda 00:00:00 UTC, domingo 23:59:59.999 UTC] da semana
* que contém `ref` (Regra 17 — semana começa à segunda).
*/
export function weekRange(ref: Date): { start: Date; end: Date; iso: string } {
const d = new Date(Date.UTC(ref.getUTCFullYear(), ref.getUTCMonth(), ref.getUTCDate()))
const dow = d.getUTCDay() // 0=Dom, 1=Seg
const diffToMonday = dow === 0 ? -6 : 1 - dow
const start = new Date(d)
start.setUTCDate(d.getUTCDate() + diffToMonday)
const end = new Date(start)
end.setUTCDate(start.getUTCDate() + 6)
end.setUTCHours(23, 59, 59, 999)
return { start, end, iso: weekIso(start) }
}
/** Semana ISO 8601 (YYYY-Www) para segunda de referência. */
export function weekIso(monday: Date): string {
// Usa algoritmo ISO: quinta da mesma semana determina o ano
const thursday = new Date(monday)
thursday.setUTCDate(monday.getUTCDate() + 3)
const year = thursday.getUTCFullYear()
const jan1 = new Date(Date.UTC(year, 0, 1))
const week = Math.floor(
((thursday.getTime() - jan1.getTime()) / 86400000 + (jan1.getUTCDay() === 0 ? 6 : jan1.getUTCDay() - 1)) / 7
) + 1
return `${year}-W${String(week).padStart(2, '0')}`
}
/** Helper: todos os session_ids no intervalo. */
function baseRows(ctx: DetectCtx) {
return ctx.db.prepare(`
SELECT session_id, project_slug, started_at, event_count, tool_calls, tools_used, skills_invoked, outcome, duration_sec
FROM sessions
WHERE started_at >= ? AND started_at <= ?
`).all(ctx.weekStartIso, ctx.weekEndIso) as Array<{
session_id: string
project_slug: string
started_at: string
event_count: number
tool_calls: number
tools_used: string
skills_invoked: string
outcome: string
duration_sec: number | null
}>
}
/** 1. Skills com taxa elevada de erro/interrupção. */
export function detectSkillsHighErrorRate(ctx: DetectCtx): Pattern[] {
const rows = baseRows(ctx)
// Agregar por skill
const bySkill = new Map<string, { total: number; fail: number; ids: string[] }>()
for (const r of rows) {
let skills: string[] = []
try { skills = JSON.parse(r.skills_invoked) } catch {}
for (const sk of skills) {
const entry = bySkill.get(sk) ?? { total: 0, fail: 0, ids: [] }
entry.total++
if (r.outcome === 'error' || r.outcome === 'interrupted') {
entry.fail++
if (entry.ids.length < 5) entry.ids.push(r.session_id)
}
bySkill.set(sk, entry)
}
}
const out: Pattern[] = []
for (const [skill, v] of bySkill) {
if (v.total < 3) continue
const ratio = v.fail / v.total
if (ratio <= 0.2) continue
const severity: Severity = ratio > 0.4 ? 'action' : 'warning'
out.push({
pattern_key: `skill_error_rate:${skill}`,
title: `Skill ${skill}: ${(ratio * 100).toFixed(0)}% das sessões falham`,
description: `De ${v.total} sessões que invocaram ${skill}, ${v.fail} terminaram em erro/interrupção.`,
severity,
metric_value: Math.round(ratio * 1000) / 1000,
sample_session_ids: v.ids,
affected_count: v.fail,
})
}
return out
}
/** 2. Tools com baixa eficiência (tool_calls/event_count elevado). */
export function detectToolsLowEfficiency(ctx: DetectCtx): Pattern[] {
const rows = baseRows(ctx)
const byTool = new Map<string, { sum: number; count: number; ids: string[] }>()
for (const r of rows) {
if (!r.event_count || r.event_count === 0) continue
const ratio = r.tool_calls / r.event_count
let tools: string[] = []
try { tools = JSON.parse(r.tools_used) } catch {}
for (const t of tools) {
const e = byTool.get(t) ?? { sum: 0, count: 0, ids: [] }
e.sum += ratio
e.count++
if (e.ids.length < 5) e.ids.push(r.session_id)
byTool.set(t, e)
}
}
const out: Pattern[] = []
for (const [tool, v] of byTool) {
if (v.count < 5) continue
const avg = v.sum / v.count
if (avg <= 0.5) continue
out.push({
pattern_key: `tool_low_efficiency:${tool}`,
title: `Tool ${tool}: rácio tool_calls/event_count médio ${avg.toFixed(2)}`,
description: `Em ${v.count} sessões, ${tool} domina o event_count. Indício de uso ineficiente ou looping.`,
severity: 'info',
metric_value: Math.round(avg * 1000) / 1000,
sample_session_ids: v.ids,
affected_count: v.count,
})
}
return out
}
/** 3. Pares (skill, tool) mais frequentes. */
export function detectSkillToolPairs(ctx: DetectCtx): Pattern[] {
const rows = baseRows(ctx)
const byPair = new Map<string, { count: number; ids: string[] }>()
for (const r of rows) {
let skills: string[] = []
let tools: string[] = []
try { skills = JSON.parse(r.skills_invoked) } catch {}
try { tools = JSON.parse(r.tools_used) } catch {}
for (const s of skills) {
for (const t of tools) {
const key = `${s}::${t}`
const e = byPair.get(key) ?? { count: 0, ids: [] }
e.count++
if (e.ids.length < 5) e.ids.push(r.session_id)
byPair.set(key, e)
}
}
}
const sorted = [...byPair.entries()].filter(([, v]) => v.count >= 5).sort((a, b) => b[1].count - a[1].count).slice(0, 5)
return sorted.map(([key, v]) => ({
pattern_key: `skill_tool_pair:${key}`,
title: `Par frequente: ${key.replace('::', ' + ')}`,
description: `Skill e tool co-ocorreram em ${v.count} sessões esta semana.`,
severity: 'info' as Severity,
metric_value: v.count,
sample_session_ids: v.ids,
affected_count: v.count,
}))
}
/** 4. Duration outliers: sessões > p95 por projecto com outcome != completed. */
export function detectDurationOutliers(ctx: DetectCtx): Pattern[] {
const rows = baseRows(ctx).filter((r) => r.duration_sec != null && r.duration_sec > 0)
const byProject = new Map<string, Array<typeof rows[number]>>()
for (const r of rows) {
const arr = byProject.get(r.project_slug) ?? []
arr.push(r)
byProject.set(r.project_slug, arr)
}
const out: Pattern[] = []
for (const [proj, arr] of byProject) {
if (arr.length < 4) continue
const durations = arr.map((r) => r.duration_sec as number).sort((a, b) => a - b)
const p95Idx = Math.max(0, Math.floor(durations.length * 0.95) - 1)
const p95 = durations[p95Idx]
const outliers = arr.filter((r) => (r.duration_sec as number) > p95 && r.outcome !== 'completed')
if (outliers.length < 3) continue
out.push({
pattern_key: `duration_outliers:${proj}`,
title: `Projecto ${proj}: ${outliers.length} sessões longas não concluídas`,
description: `Sessões com duração acima do p95 (${p95}s) e outcome != completed. Sinal de sessões penduradas.`,
severity: 'warning',
metric_value: p95,
sample_session_ids: outliers.slice(0, 5).map((r) => r.session_id),
affected_count: outliers.length,
})
}
return out
}
/** 5. Sessões abandonadas (event_count < 3 AND outcome=unknown). */
export function detectAbandonedSessions(ctx: DetectCtx): Pattern[] {
const rows = ctx.db.prepare(`
SELECT session_id FROM sessions
WHERE started_at >= ? AND started_at <= ?
AND event_count < 3 AND outcome = 'unknown'
`).all(ctx.weekStartIso, ctx.weekEndIso) as Array<{ session_id: string }>
if (rows.length < 5) return []
return [{
pattern_key: 'abandoned_sessions',
title: `${rows.length} sessões abandonadas esta semana`,
description: `Sessões com menos de 3 eventos e outcome=unknown — tipicamente abertas e descartadas.`,
severity: 'info',
metric_value: rows.length,
sample_session_ids: rows.slice(0, 5).map((r) => r.session_id),
affected_count: rows.length,
}]
}
/** 6. Crescimento de complexidade: avg(tool_calls) actual vs semana anterior. */
export function detectGrowingComplexity(ctx: DetectCtx, prevWeekStartIso: string, prevWeekEndIso: string): Pattern[] {
const curRows = baseRows(ctx)
const prevRows = ctx.db.prepare(`
SELECT skills_invoked, tool_calls FROM sessions
WHERE started_at >= ? AND started_at <= ?
`).all(prevWeekStartIso, prevWeekEndIso) as Array<{ skills_invoked: string; tool_calls: number }>
const curBySkill = new Map<string, { sum: number; count: number; ids: string[] }>()
for (const r of curRows) {
let sk: string[] = []
try { sk = JSON.parse(r.skills_invoked) } catch {}
for (const s of sk) {
const e = curBySkill.get(s) ?? { sum: 0, count: 0, ids: [] }
e.sum += r.tool_calls
e.count++
if (e.ids.length < 5) e.ids.push(r.session_id)
curBySkill.set(s, e)
}
}
const prevBySkill = new Map<string, { sum: number; count: number }>()
for (const r of prevRows) {
let sk: string[] = []
try { sk = JSON.parse(r.skills_invoked) } catch {}
for (const s of sk) {
const e = prevBySkill.get(s) ?? { sum: 0, count: 0 }
e.sum += r.tool_calls
e.count++
prevBySkill.set(s, e)
}
}
const out: Pattern[] = []
for (const [skill, cur] of curBySkill) {
if (cur.count < 5) continue
const curAvg = cur.sum / cur.count
const prev = prevBySkill.get(skill)
if (!prev || prev.count < 3) continue
const prevAvg = prev.sum / prev.count
if (prevAvg === 0 || curAvg <= prevAvg * 1.3) continue
out.push({
pattern_key: `growing_complexity:${skill}`,
title: `Skill ${skill}: tool_calls médio +${Math.round((curAvg / prevAvg - 1) * 100)}% vs semana anterior`,
description: `Média de tool_calls/sessão subiu de ${prevAvg.toFixed(1)} para ${curAvg.toFixed(1)}.`,
severity: 'warning',
metric_value: Math.round(curAvg * 10) / 10,
sample_session_ids: cur.ids,
affected_count: cur.count,
})
}
return out
}
/** Orquestra todos os detectores para a semana indicada. */
export function detectPatterns(
dbWrapper: SessionsDb,
weekStart: Date,
weekEnd: Date,
): Pattern[] {
const db = dbWrapper.rawDb()
const ctx: DetectCtx = {
db,
weekStartIso: iso(weekStart),
weekEndIso: iso(weekEnd),
}
const prevStart = new Date(weekStart); prevStart.setUTCDate(prevStart.getUTCDate() - 7)
const prevEnd = new Date(weekEnd); prevEnd.setUTCDate(prevEnd.getUTCDate() - 7)
return [
...detectSkillsHighErrorRate(ctx),
...detectToolsLowEfficiency(ctx),
...detectSkillToolPairs(ctx),
...detectDurationOutliers(ctx),
...detectAbandonedSessions(ctx),
...detectGrowingComplexity(ctx, iso(prevStart), iso(prevEnd)),
]
}
/** Converte Pattern + contexto em PatternRecord pronto a persistir. */
export function toPatternRecord(p: Pattern, weekIso: string, consecutiveWeeks: number): PatternRecord {
return {
detected_at: new Date().toISOString(),
week_iso: weekIso,
pattern_key: p.pattern_key,
title: p.title,
description: p.description,
severity: p.severity,
metric_value: p.metric_value,
sample_session_ids: p.sample_session_ids,
affected_count: p.affected_count,
consecutive_weeks: consecutiveWeeks,
}
}