mirror of
https://github.com/proffesor-for-testing/agentic-qe.git
synced 2026-09-19 08:45:47 +08:00
aqe dream cycles
This commit is contained in:
Profa
@@ -0,0 +1,290 @@
|
||||
# AQE Dream Cycles & Neural Learning — Full Flow Analysis
|
||||
|
||||
## Overview
|
||||
|
||||
The system uses a **biological sleep-consolidation metaphor**: during normal QE work, experiences are captured into patterns; periodically, a "dream cycle" runs spreading activation across a concept graph to discover cross-domain insights that improve future agent recommendations.
|
||||
|
||||
---
|
||||
|
||||
## Phase 1: Pattern Collection (Happens Automatically)
|
||||
|
||||
Every time a user runs a QE task — via **agents/skills**, **MCP tools**, or **CLI** — hook events fire on the EventBus:
|
||||
|
||||
| User Action | Event Fired | What Gets Captured |
|
||||
|---|---|---|
|
||||
| Generate tests | `qe:post-test-generation` | Test strategy, framework, success rate |
|
||||
| Run coverage analysis | `qe:coverage-gap-identified` | Gap patterns, risk scores |
|
||||
| Execute quality gate | `qe:quality-score` | Quality metrics, domain guidance |
|
||||
| Any agent completes | `qe:agent-completion` | Duration, outcome, confidence |
|
||||
|
||||
The `QEReasoningBank` stores these as patterns in `qe_patterns` (SQLite) with:
|
||||
- 384-dimensional vector embeddings (all-MiniLM-L6-v2 via ONNX) for HNSW search
|
||||
- Confidence scores updated via **asymmetric learning** (success: +0.1, failure: -1.0 — a 10:1 penalty ratio)
|
||||
- Tier classification: `short-term` → promoted to `long-term` after 3+ successful uses
|
||||
|
||||
**Key files:** `v3/src/learning/qe-hooks.ts`, `v3/src/learning/qe-reasoning-bank.ts`, `v3/src/learning/experience-capture.ts`
|
||||
|
||||
---
|
||||
|
||||
## Phase 2: Dream Cycle Triggers
|
||||
|
||||
The `DreamScheduler` (`v3/src/learning/dream/dream-scheduler.ts`) triggers dreams via four mechanisms:
|
||||
|
||||
| Trigger | Default Threshold |
|
||||
|---|---|
|
||||
| **Time-based** | Every 60 minutes |
|
||||
| **Experience buffer** | Every 20 captured experiences |
|
||||
| **Quality gate failure** | Any failed quality gate |
|
||||
| **Background worker** | `LearningConsolidationWorker` every 30 min |
|
||||
|
||||
Minimum gap between dreams: **5 minutes**. Users can also trigger manually:
|
||||
|
||||
```bash
|
||||
# CLI
|
||||
aqe learning consolidate
|
||||
|
||||
# MCP tool
|
||||
qe/learning/dream { action: "dream" }
|
||||
|
||||
# Session-end hook (automatic)
|
||||
.claude/helpers/learning-hooks.sh session-end
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Phase 3: What Happens During a Dream
|
||||
|
||||
**Core engine:** `v3/src/learning/dream/dream-engine.ts`
|
||||
|
||||
```
|
||||
1. LOAD CONCEPTS
|
||||
└─ Top 200 patterns (confidence >= 0.3) → concept_nodes + concept_edges
|
||||
|
||||
2. BUILD GRAPH IN MEMORY
|
||||
└─ Pattern nodes, domain nodes, similarity edges, co-occurrence edges
|
||||
|
||||
3. SPREADING ACTIVATION (the actual "dreaming")
|
||||
└─ Loop until timeout (5s quick / 10s default / 30s full):
|
||||
→ Pick random concept node
|
||||
→ Inject noise (simulates REM sleep's reduced logical filtering)
|
||||
→ Propagate activation to neighbors: weight × 0.5 spread factor
|
||||
→ Decay all activations by 10% per iteration
|
||||
→ Track co-activation pairs
|
||||
→ Detect novel associations (both activated, no existing strong edge)
|
||||
|
||||
4. GENERATE INSIGHTS (4 detectors)
|
||||
├─ Pattern merges: Redundant patterns to consolidate
|
||||
├─ Novel associations: Cross-domain connections discovered
|
||||
├─ Optimizations: Low-performing patterns to review
|
||||
└─ Gap detection: Errors without resolution patterns
|
||||
|
||||
5. PERSIST
|
||||
└─ Insights → dream_insights table (filtered: novelty >= 0.3, max 10)
|
||||
└─ Cycle metadata → dream_cycles table
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Phase 4: Applying Insights
|
||||
|
||||
Insights sit in `dream_insights` as pending until applied:
|
||||
|
||||
```bash
|
||||
# View pending insights
|
||||
MCP: qe/learning/dream { action: "insights" }
|
||||
|
||||
# Apply a specific insight (creates a real QE pattern)
|
||||
MCP: qe/learning/dream { action: "apply", insightId: "..." }
|
||||
```
|
||||
|
||||
When applied, an insight becomes a new pattern in `qe_patterns` with type mapping:
|
||||
- `cross-domain` → `coverage-strategy`
|
||||
- `novel-path` → `test-template`
|
||||
- `cluster` → `refactor-safe`
|
||||
- `high-activation` → `assertion-pattern`
|
||||
|
||||
If `autoApplyHighConfidenceInsights` is enabled, insights with confidence >= 0.8 are applied automatically.
|
||||
|
||||
---
|
||||
|
||||
## Phase 5: Learned Patterns Influence Future Work
|
||||
|
||||
On every new task, `QEReasoningBank.routeTask()` runs:
|
||||
|
||||
1. Embeds the task description (384-dim)
|
||||
2. HNSW search in `qe_patterns` for top-k similar patterns (O(log n))
|
||||
3. Scores: `similarity × 0.3 + performance × 0.4 + capabilities × 0.3`
|
||||
4. Returns `recommendedAgent` + `guidance[]` array
|
||||
|
||||
This is visible in the session-start hook output:
|
||||
```json
|
||||
{
|
||||
"recommendedAgent": "qe-test-architect",
|
||||
"confidence": 0.313,
|
||||
"guidance": ["Follow AAA pattern...", "One assertion per test..."],
|
||||
"patternCount": 2
|
||||
}
|
||||
```
|
||||
|
||||
Dream-generated patterns appear in these search results, meaning **the system's own dreaming directly shapes what guidance future agents receive**.
|
||||
|
||||
---
|
||||
|
||||
## Complete Flow Diagram
|
||||
|
||||
```
|
||||
USER RUNS QE TASK (skill, agent, MCP tool, or CLI)
|
||||
│
|
||||
▼
|
||||
┌─── PRE-HOOK ───────────────────────────────┐
|
||||
│ routeTask() → HNSW search qe_patterns │
|
||||
│ Returns: agent recommendation + guidance[] │
|
||||
│ (includes dream-generated patterns!) │
|
||||
└────────────────────────────────────────────┘
|
||||
│
|
||||
▼
|
||||
[ Agent executes with injected guidance ]
|
||||
│
|
||||
▼
|
||||
┌─── POST-HOOK ──────────────────────────────┐
|
||||
│ recordOutcome() → asymmetric confidence │
|
||||
│ storePattern() if quality >= 0.7 │
|
||||
│ promote() if usage >= 3 │
|
||||
│ DreamScheduler.recordExperience() │
|
||||
└────────────────────────────────────────────┘
|
||||
│
|
||||
▼
|
||||
┌─── DREAM TRIGGER ─────────────────────────┐
|
||||
│ 20 experiences / 60 min / gate failure │
|
||||
└────────────────────────────────────────────┘
|
||||
│
|
||||
▼
|
||||
┌─── DREAM CYCLE ───────────────────────────┐
|
||||
│ Load patterns → Build concept graph │
|
||||
│ Spreading activation with noise │
|
||||
│ Detect: merges, novel links, gaps, opts │
|
||||
│ Store insights → dream_insights │
|
||||
└────────────────────────────────────────────┘
|
||||
│
|
||||
▼
|
||||
┌─── APPLY INSIGHT ─────────────────────────┐
|
||||
│ Insight → new qe_pattern with embedding │
|
||||
│ Available in HNSW for next routeTask() │
|
||||
└────────────────────────────────────────────┘
|
||||
│
|
||||
└──────────► NEXT TASK (loop back to top)
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Pattern Lifecycle Management
|
||||
|
||||
Beyond dreams, patterns have ongoing lifecycle management (`v3/src/learning/pattern-lifecycle.ts`):
|
||||
|
||||
- **Daily confidence decay**: -0.01/day for unused patterns
|
||||
- **Deprecation**: 3+ consecutive failures → deprecated
|
||||
- **Staleness**: 30 days without use → marked stale
|
||||
- **Quarantine**: Below 0.3 confidence → needs 10 consecutive successes to rehabilitate
|
||||
|
||||
---
|
||||
|
||||
## Key Files Reference
|
||||
|
||||
| Component | File |
|
||||
|---|---|
|
||||
| DreamScheduler | `v3/src/learning/dream/dream-scheduler.ts` |
|
||||
| DreamEngine | `v3/src/learning/dream/dream-engine.ts` |
|
||||
| ConceptGraph | `v3/src/learning/dream/concept-graph.ts` |
|
||||
| SpreadingActivation | `v3/src/learning/dream/spreading-activation.ts` |
|
||||
| InsightGenerator | `v3/src/learning/dream/insight-generator.ts` |
|
||||
| Dream types | `v3/src/learning/dream/types.ts` |
|
||||
| DB schema | `v3/src/learning/dream/schema/dream-tables.sql` |
|
||||
| QE hooks (capture) | `v3/src/learning/qe-hooks.ts` |
|
||||
| Experience capture | `v3/src/learning/experience-capture.ts` |
|
||||
| ReasoningBank | `v3/src/learning/qe-reasoning-bank.ts` |
|
||||
| Pattern store | `v3/src/learning/pattern-store.ts` |
|
||||
| Asymmetric learning | `v3/src/learning/asymmetric-learning.ts` |
|
||||
| Pattern lifecycle | `v3/src/learning/pattern-lifecycle.ts` |
|
||||
| Pattern evolution | `v3/src/integrations/agentic-flow/reasoning-bank/pattern-evolution.ts` |
|
||||
| MCP dream tool | `v3/src/mcp/tools/learning-optimization/dream.ts` |
|
||||
| Learning CLI | `v3/src/cli/commands/learning.ts` |
|
||||
| Background worker | `v3/src/workers/workers/learning-consolidation.ts` |
|
||||
| Shell hooks | `.claude/helpers/learning-hooks.sh` |
|
||||
|
||||
---
|
||||
|
||||
## Access Points Summary
|
||||
|
||||
| Method | Command/Tool | What It Does |
|
||||
|---|---|---|
|
||||
| **CLI** | `aqe learning stats` | View pattern counts, success rates |
|
||||
| **CLI** | `aqe learning consolidate` | Promote, prune, run dream cycle |
|
||||
| **CLI** | `aqe learning daemon --interval 3600` | Background consolidation loop |
|
||||
| **MCP** | `qe/learning/dream {action:"dream"}` | Run dream cycle on demand |
|
||||
| **MCP** | `qe/learning/dream {action:"insights"}` | View pending insights |
|
||||
| **MCP** | `qe/learning/dream {action:"apply"}` | Apply insight as real pattern |
|
||||
| **MCP** | `qe/learning/dream {action:"status"}` | Dream scheduler status |
|
||||
| **Hook** | `session-end` | Auto-consolidate on session close |
|
||||
| **Hook** | `session-start` | Load patterns, show recommendations |
|
||||
| **Automatic** | `DreamScheduler` | Timer/threshold/event triggers |
|
||||
|
||||
---
|
||||
|
||||
## Batch Apply Results (2026-02-22)
|
||||
|
||||
A batch apply of all pending actionable dream insights was performed:
|
||||
|
||||
### Before vs After
|
||||
|
||||
| Metric | Before | After |
|
||||
|---|---|---|
|
||||
| **Total patterns in DB** | 6,029 | 6,644 |
|
||||
| **Dream-generated patterns** | 0 | 615 |
|
||||
| **Pending insights** | 615 | 0 |
|
||||
| **Applied insights** | 5 | 620 |
|
||||
| **DB integrity** | ok | ok |
|
||||
|
||||
### Dream System State
|
||||
|
||||
- **100 total dream cycles** have been executed historically
|
||||
- **0 pending insights** remaining — all actionable insights applied
|
||||
- All 615 new patterns are type `coverage-strategy` (from `novel_association` insights)
|
||||
- Patterns entered as `short-term` tier with confidence ~0.58
|
||||
- Patterns need vector embeddings (generated lazily on first HNSW search) and successful uses to promote to `long-term`
|
||||
|
||||
### Batch Apply Method
|
||||
|
||||
For bulk insight application, direct SQL is far more efficient than per-insight CLI calls (which reinitialize the engine each time). The SQL approach:
|
||||
|
||||
```sql
|
||||
-- Create patterns from unapplied actionable insights
|
||||
INSERT INTO qe_patterns (id, pattern_type, qe_domain, domain, name, description, confidence, tier, template_json, context_json, created_at, updated_at)
|
||||
SELECT
|
||||
'dream-' || di.id,
|
||||
CASE di.insight_type
|
||||
WHEN 'novel_association' THEN 'coverage-strategy'
|
||||
WHEN 'cross-domain' THEN 'coverage-strategy'
|
||||
WHEN 'pattern_merge' THEN 'refactor-safe'
|
||||
WHEN 'optimization' THEN 'test-template'
|
||||
WHEN 'gap_detection' THEN 'coverage-strategy'
|
||||
ELSE 'test-template'
|
||||
END,
|
||||
'learning-optimization', 'learning-optimization',
|
||||
'Dream: ' || di.insight_type || ' - ' || substr(di.id, -8),
|
||||
substr(di.description, 1, 500),
|
||||
di.confidence_score, 'short-term',
|
||||
json_object('type', 'workflow', 'content', COALESCE(di.suggested_action, di.description), 'variables', json_array()),
|
||||
json_object('tags', json_array('dream-generated', di.insight_type), 'complexity', 'medium'),
|
||||
datetime('now'), datetime('now')
|
||||
FROM dream_insights di
|
||||
WHERE di.actionable = 1 AND di.applied = 0;
|
||||
|
||||
-- Mark all as applied
|
||||
UPDATE dream_insights SET applied = 1, pattern_id = 'dream-' || id
|
||||
WHERE actionable = 1 AND applied = 0;
|
||||
```
|
||||
|
||||
**Important:** Always backup the database before batch operations:
|
||||
```bash
|
||||
cp .agentic-qe/memory.db .agentic-qe/memory.db.bak-$(date +%s)
|
||||
```
|
||||
Reference in New Issue
Block a user