Lalitkumar Bhamare
13 KiB
Agentic QE v3 Framework — Technical & Architectural Glossary
Core Architectural Patterns
DDD (Domain-Driven Design) — The framework is structured around 12 bounded contexts (test-generation, test-execution, coverage-analysis, quality-assessment, defect-intelligence, requirements-validation, code-intelligence, security-compliance, contract-testing, visual-accessibility, chaos-resilience, learning-optimization). Each domain owns its agents, events, and data. Prevents cross-domain coupling so teams can evolve domains independently. (ADR-001)
ADR (Architecture Decision Records) — 57 formal design decisions (ADR-001 through ADR-057 + MADR-001) that document why each architectural choice was made. Each ADR includes context, decision, consequences, and status. They serve as the institutional memory of the framework's evolution.
QCSD (Quality Conscious Software Delivery) — The 5-phase swarm model that structures quality across the entire SDLC: Ideation (planning) → Refinement (story prep) → Development (coding) → CI/CD Verification (release) → Production (telemetry) . Each phase produces a binary-style decision (GO/NO-GO, READY/NOT-READY, SHIP/HOLD, RELEASE/BLOCK) with cross-phase signal propagation.
SPARC (Specification, Pseudocode, Architecture, Refinement, Completion) — A methodology for structured agent work. Breaks complex tasks into 5 phases so agents don't skip steps. The sparc-orchestrator agent coordinates this cycle.
PACT (Proactive, Autonomous, Collaborative, Targeted) — The foundational principles of Agentic QE. Agents are proactive (find problems before asked), autonomous (self-coordinate), collaborative (share knowledge), and targeted (focus on highest-risk areas).
Testing Frameworks & Methodologies
HTSM (Holistic Testing Strategy Mapping) v6.3 — James Bach's framework for comprehensive test strategy. Evaluates 6 quality dimensions: Coverage, Reliability, Performance, Security, Accessibility, Maintainability. Used by the Ideation swarm during PI/Sprint planning.
SFDIPOT (Scope, Feasibility, Dependencies, Integration, Performance, Operations, Testability) — Product factors analysis model used during Refinement. The qe-product-factors-assessor agent evaluates each dimension to determine story readiness.
London/Chicago TDD — Two schools of Test-Driven Development. London (mockist) isolates units with mocks for fast feedback. Chicago (classicist) tests real collaborators for integration confidence. The qe-tdd-specialist supports both.
Mutation Testing — Injects small code changes (mutants) into production code and checks if tests catch them. A mutant that survives means the test suite has a gap. Measures test quality, not just coverage.
Chaos Engineering — Controlled fault injection in production-like environments to validate resilience. Tests what happens when services fail, networks partition, or resources exhaust.
Algorithms & Data Structures
HNSW (Hierarchical Navigable Small World) — O(log n) approximate nearest-neighbour vector search algorithm. Powers semantic search across patterns, test cases, and code embeddings. Replaces brute-force O(n) scanning for coverage gap detection. (ADR-003)
MinCut — Graph theory algorithm that finds the minimum number of edges whose removal disconnects a graph. Used by the Strange Loop self-awareness system to detect bottlenecks and single points of failure in the agent swarm. Runs in ~30μs. (ADR-047)
TinyDancer — The lightweight intelligent model router (TD-003, ADR-026). Classifies task complexity and routes to the optimal Claude model tier: Haiku for simple tasks, Sonnet for medium, Opus for complex. Uses confidence scoring — if uncertain, triggers multi-model verification or human review. Named for making "efficient, graceful routing decisions."
PageRank — Google's link analysis algorithm adapted for code importance ranking. Identifies which modules/functions have the highest "influence" in the codebase to prioritise testing effort.
Floyd-Warshall — All-pairs shortest path algorithm used in causal discovery to trace root cause chains across service dependencies.
Tarjan's Algorithm — Detects strongly connected components (circular dependencies) in dependency graphs. Used by qe-dependency-mapper.
CRDT (Conflict-free Replicated Data Types) — Data structures that allow concurrent updates from multiple agents without coordination locks. Ensures eventual consistency across distributed swarm state.
Neural & ML Architectures
ReasoningBank — Pattern storage and retrieval system with transformer embeddings. Stores successful QE patterns (test strategies, defect fixes, routing decisions) indexed for semantic search. Achieves 114k records/sec throughput. Enables experience replay and cross-agent knowledge transfer. (ADR-021)
RuVector — The primary neural backbone combining Q-Learning persistence, SONA self-optimisation, and hypergraph code intelligence. Provides AST-level code understanding with vector embeddings. (ADR-017, ADR-050)
SONA (Self-Optimizing Neural Architecture) — LoRA-style fine-tuning with EWC++ memory preservation. Agents improve over time without catastrophic forgetting of previous patterns.
EWC++ (Elastic Weight Consolidation) — Prevents catastrophic forgetting when neural models learn new tasks. Protects important weights from being overwritten, so agents retain old patterns while learning new ones.
Flash Attention — Memory-efficient attention mechanism that reduces the O(n²) cost of standard attention to near-linear. Used for processing large codebases without running out of context. Target: 2.49x–7.47x speedup. (ADR-051)
GNN (Graph Neural Networks) — Neural networks that operate on graph-structured data. Used for code embedding — understands relationships between functions, classes, and modules rather than treating code as flat text.
Q-Learning Router — Reinforcement learning for agent routing. Maintains Q-value tables that learn which agent types succeed on which task types. Converges to optimal routing without explicit programming. (ADR-022)
GOAP (Goal-Oriented Action Planning) — Gaming AI technique adapted for QE. Given a goal state (e.g., "coverage > 80%"), discovers action sequences to achieve it by searching through available operations. 52 available actions in the v3 planner. (ADR-046)
Coherence & Formal Verification (Prime Radiant)
Prime Radiant — The mathematical coherence engine suite with 6 WASM-accelerated engines:
-
Sheaf Cohomology — Detects contradictions across agent belief spaces. If agent A says "test passes" and agent B says "test fails", sheaf analysis finds the inconsistent patch.
-
Spectral Analysis — Eigenvalue decomposition of the agent Laplacian matrix. Predicts state-space collapse (when the swarm is about to lose coherence) before it happens.
-
Causal Inference — Distinguishes correlation from causation. When test failures spike alongside a deploy, determines whether the deploy caused the failures or if it's coincidental.
-
Category Theory — Verifies type-safe mappings between domains. Functors and natural transformations ensure domain operations compose correctly.
-
Homotopy Type Theory — Formal verification of pipeline correctness. Proves that two pipeline paths produce equivalent results (path equality).
-
Witness/Audit Engine — Records every decision with a cryptographic witness for deterministic replay and debugging.
Lambda (λ) Coherence — The unified coherence signal derived from MinCut analysis. Drives the 4-tier compute allocation: Reflex (<1ms, low energy), Retrieval (~10ms, cached patterns), Escalate (full reasoning), Human (manual override). (ADR-030)
Self-Organising & Biological Patterns
Strange Loop — Self-awareness pattern where the swarm observes itself, builds a self-model, detects anomalies, and heals — an Observe→Model→Decide→Act cycle. Enables self-healing without human intervention. Response time ~12s. (ADR-031)
Time Crystal Scheduling (Kuramoto CPG) — Borrowed from coupled oscillator physics. Agents self-organise into test phases (Unit→Integration→E2E→Performance) without a central scheduler. Winner-take-all dynamics determine which phase is active. (ADR-032)
Morphogenetic Networks — Pattern formation without centralised control, inspired by biological development. Agents form structure through local interactions only, enabling scalability to arbitrary swarm sizes.
Early Exit Testing — Lambda-stability detection lets the system skip remaining analysis when confidence is already high. Saves 30–50% compute on high-confidence runs. (ADR-033)
STDP (Spike-Timing Dependent Plasticity) — Biological learning rule adapted for causal discovery. If event A consistently precedes event B, the causal link A→B is strengthened. Used for root cause analysis. (ADR-035)
Infrastructure & Self-Healing
Infrastructure Self-Healing (ADR-057) — Extends Strange Loop to infrastructure. Components:
- TestOutputObserver — Parses stderr for infrastructure error signatures (SAP RFC, Salesforce, payment gateway, etc.)
- RecoveryPlaybook — YAML definitions of per-service health checks, recovery commands, and verification steps
- InfraActionExecutor — Health check → recover → exponential backoff → verify cycle
- 34+ built-in signatures — Pre-configured patterns for common infrastructure failures
AgentDB — The unified persistence layer replacing 6+ separate memory systems. SQLite backend with HNSW vector index, achieving 150x–12,500x search improvement over brute force. (ADR-009, ADR-038)
Communication & Protocols
A2A (Agent-to-Agent Protocol) — JSON-RPC 2.0 based protocol for inter-agent communication. 68 agents publish Agent Cards describing their capabilities. Discovery endpoint for dynamic agent lookup. (ADR-054)
AG-UI (Agent-to-UI Protocol) — SSE-based real-time event streaming with 19 event types. 100ms P95 latency for live UI updates during swarm execution. (ADR-053)
QUIC Swarm — Low-latency multiplexed transport for agent coordination. Target <10ms inter-agent latency with built-in congestion control. (ADR-051)
7 Coordination Protocols — Morning-Sync, Quality-Gate, Regression-Prevention, Coverage-Driven, TDD-Cycle, Security-Audit, Learning-Consolidation. Each defines a cross-domain event choreography. (ADR-002)
MCP (Model Context Protocol) — The tool interface that exposes 100+ QE operations as callable functions. Agents and external tools interact with the framework through MCP endpoints. (ADR-010)
Routing & Cost Optimisation
3-Tier Model Routing (ADR-026) — Tasks are classified by complexity and routed to the cheapest capable model:
- Tier 1 — Agent Booster (<1ms, $0): Rust/WASM mechanical transforms (var→const, add-types, remove-console). 352x faster than LLM calls.
- Tier 2 — Haiku (~500ms): Simple bug fixes, low complexity tasks.
- Tier 3 — Sonnet/Opus (2–5s): Architecture, security, complex reasoning.
Result: 87% cost reduction via multi-model routing. (ADR-051)
Security
Zero-Trust Architecture — Every agent interaction is verified. No implicit trust between agents.
Claims-Based Authorization (ADR-010) — Fine-grained access control across swarm agents and MCP tools.
OAuth2.1 + Rate Limiting (ADR-012) — MCP security layer preventing abuse and ensuring authenticated access.
OSV Client (ADR-013) — Automated CVE scanning against the Open Source Vulnerability database.
Skill Validation
Deterministic Skill Validation (ADR-056) — 5-tier trust system for skills:
- Tier 0: Advisory (SKILL.md only)
- Tier 1: Structured (+ JSON Schema)
- Tier 2: Validated (+ executable validator)
- Tier 3: Verified (+ 5+ evaluation cases) — 46 skills at this tier
- Tier 4: Certified (cross-model validation, 95%+ pass rate)
Key Performance Numbers
| Metric | Value |
|---|---|
| ADRs implemented | 54 |
| DDD bounded contexts | 12 |
| QE agents | 47+ |
| Invocable skills | 46 (Tier 3) |
| ReasoningBank throughput | 114k records/sec |
| Agent routing P95 latency | 62ms |
| MinCut calculation | ~30μs |
| Coverage gap detection | <100ms on 500 files |
| Self-healing response | ~12s |
| Strange Loop cycle | ~45ms |
| Coherence engines | 6 (WASM-accelerated) |
| Infrastructure signatures | 34+ built-in |