# CHANGELOG

## v3.0.0a1 (2026-05-01)

### Added — Track A: Foundation
- **Canonical Event Runtime** (`purpose_agent/runtime/`)
  - `PAEvent` frozen dataclass with 30 event kinds
  - `EventBus` async pub/sub with backpressure, replay, lane isolation
  - `Visibility` levels (public/internal/debug) — no raw CoT leaks
  - `streaming_v3.py` AG-UI compatible stream adapters
- **Durable Execution** (`runtime/checkpoint.py`, `runtime/state.py`)
  - `RunState` with node tracking, idempotency keys, serialization
  - `Checkpointer` protocol + InMemory/JSONL/SQLite implementations
  - Resume from crash, HITL pause/resume
- **Memory Homeostasis** (`memory_homeostasis.py`)
  - `MemoryBudget` hard limits (active cards, injected tokens, per-kind caps)
  - `MemoryArchive` append-only cold storage
  - `ConsolidationEngine` (cluster → merge → compress → hibernate)
  - `QFunctionRetriever` multi-signal ranking under token budget

### Added — Track B: Interoperability
- **Protocol Stack** (`purpose_agent/protocols/`)
  - `MCPToolBridge` — MCP server registration, tool wrapping, security policy
  - `A2AClient` + `AgentCard` — agent-to-agent delegation with circuit breaker
  - `AGUIAdapter` — PAEvent → AG-UI frontend streaming
  - `parse_agents_md()` — repo-local instruction parser with precedence
- **Quorum Coordinator** (`quorum.py`)
  - Agreement → merge, disagreement → escalate, critical risk → HITL
  - `CriticEnsemble` (correctness/safety/efficiency/alignment)

### Added — Track C: Intelligence
- **LLM Call Router** (`routing.py`)
  - `TaskComplexityClassifier` (simple/moderate/complex/critical)
  - `ModelSelector` local-first, smallest-capable
  - Cost budget hard cap
- **MAS Generator** (`mas_generator.py`)
  - `generate(use_case)` → complete system (agents + flow + tools + evals + routing)
  - 5 templates: code, security, research, data, operations
- **Skill Genome** (`purpose_agent/skills/`)
  - `SkillCard` versioned, testable, evolvable
  - `SkillGenome` lineage tracking + rollback
  - `SkillCI` immune scan → fitness test → promote/archive + mutation

### Added — Track D: Optimization
- **Capability Fingerprint** (`optimization/fingerprint.py`)
  - Domains, tool motifs, action patterns, failure modes, latency/cost
- **Dataset Builder** (`optimization/dataset.py`)
  - Trace → filtered examples with immune scan + dedup + train/val/test split
- **Prompt Pack** (`optimization/prompt_pack.py`)
  - Token-budget-compliant optimization artifact (epigenetics first)
- **Shadow Evaluator** (`optimization/shadow_eval.py`)
  - Candidate vs baseline comparison, promote/rollback decisions
- **Agentic Optimizer** (`optimization/optimizer.py`)
  - Trend monitoring: improving → continue, plateau → optimize, degrading → rollback
- **Distillation Planning** (`optimization/lora_plan.py`)
  - Dry-run plans (none/prompt_pack/lora/distill/pdq), no GPU required for planning

### Added — Track E: Hardening
- **Robustness Benchmark** (`benchmark_v3.py`)
  - 35+ checks: core loop, tools, immune, memory, routing, skills, events, protocols
  - Runs with mock backend in CI

### Backward Compatibility
- All v2.1.1 public imports still work
- `pa.purpose()`, `Team.run()`, `Spark`, `Flow`, `swarm`, `Council`, `Vault` unchanged
- Old names (`Agent`, `Graph`, `parallel`, `Conversation`, `KnowledgeStore`) are aliases
- `tests/compat/test_public_api_211.py` validates full API surface

---

## v2.1.1 (2026-04-30)
- Creative naming: Spark, Flow, swarm, Council, Vault
- 19/19 production test pass with Llama-3.3-70B
- Published on PyPI

## v2.0.0 (2026-04-29)
- V2 kernel: evidence-gated memory, immune system, RunMode, trace, compiler
- 13 research papers implemented
- 6 breakthroughs: self-improving critic, MoH, hindsight, evolution, adversarial
- TOML prompts, robust_parser, secure tools
- 119/119 tests pass

## v0.2.0 (2026-04-28)
- Initial release: Actor, Purpose Function, Experience Replay, Optimizer
- SLM-native backends, streaming, tools, multi-agent, HITL, evaluation
- Published on PyPI