nyxCore-Systems · SimplyLiz · Apr 10, 2026 · Apr 9, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -0,0 +1,149 @@
+# Changelog
+
+All notable changes to SwarmWire are documented here.
+
+---
+
+## [1.5.0] — 2026-04-09
+
+### New Features
+
+**Execution**
+- `reduceTrajectory` — AgentDiet-style trajectory pruning (drop empty/duplicate/superseded tool results, token budget trim). 39-60% input token reduction.
+- `SpeculativeToolExecutor` — PASTE-inspired prefetch of likely tool calls in parallel while the LLM generates.
+- `createReducedSkillSet` / `selectRelevantTools` — Progressive skill disclosure: compact one-liners first, full schemas on demand. ~48% prompt compression.
+
+**Memory**
+- `AMem` — A-MEM Zettelkasten living memory graph. On every write, notes auto-link to related memories via cosine similarity.
+- `TemporalMemory` — CMA temporal decay + spreading activation. Strength decays per-hour, reinforces on access, propagates relevance to temporal neighbors.
+- `SelfEditingMemory` — Letta/MemGPT-style named memory blocks. Agents read and mutate versioned text blocks mid-execution. Full edit history + revert.
+- `createFlatVectorStore` / `createPineconeStore` / `createQdrantStore` / `createRedisVectorStore` — External vector store adapters, all implementing `MemoryBackend`.
+- `SleepTimeAgent` — LLM-driven background consolidation. Synthesizes insights from recent memory during idle periods.
+
+**Core**
+- `ReputationBoard` — `MessageBoard` extended with per-agent reputation scoring. Upvotes, citations, correct answers drive scores. Findings weighted by sender reputation.
+- Typed DI — `AgentContext<TDeps>` and `AgentDefinition<TInput, TOutput, TDeps>`. Agents declare typed dependencies; `context.deps` is fully typed at callsite.
+
+**Testing & Evaluation**
+- `evalTrajectory` / `compareTrajectories` — TRACE-style multi-dimension trajectory evaluation: step efficiency, tool precision, backtrack rate, plan adherence, outcome quality.
+
+**Workflow**
+- `StateMachine` / `buildLinearStateMachine` — LangGraph-style directed graph with cycles, conditional edges, and `maxIterations` guard.
+
+**Patterns**
+- `runLoop` — LoopAgent primitive. Runs an agent iteratively until convergence (`shouldStop` predicate, DONE signal, or `maxIterations`). Full iteration history.
+
+**Session**
+- `BranchManager` — Fork a session at any message index to explore alternative continuations. Diff, merge, and tree visualization.
+
+**Observability**
+- `exportToOTLP` / `createOTelExporter` / `withOTelExport` — Auto-push traces to any OTLP/HTTP endpoint (Jaeger, Tempo, Honeycomb, OTEL Collector) after execution.
+
+---
+
+## [1.4.0] — 2026-04-08
+
+### New Features
+
+**Execution**
+- `TimeTravelStore` — Rewind to any step and fork execution from that point with optional step modifications.
+- `RollbackManager` — Snapshot state before tool calls; undo individual or full-execution actions in reverse order.
+
+**Optimizer**
+- `PromptOptimizer` — DSPy-style prompt optimization. Bootstraps few-shot examples from `DistillationCollector`, generates prompt variants via LLM, scores against training pairs.
+
+**Testing**
+- `EvalHarness` — Named harnesses with run history, pass-rate tracking, and regression detection. Computes trend (`improving` / `stable` / `degrading`) from last 3 runs.
+
+**Tools**
+- `createNodeSandbox` / `createDockerSandbox` / `createE2BSandbox` — Code execution sandbox with three backends. Returns a `Tool` for `agent.tools[]`.
+- `createBrowserTool` — Playwright-backed browser automation tool (navigate, click, type, screenshot, extract).
+- `createComputerUseTool` — Anthropic Computer Use API tool wrapper.
+
+**Patterns**
+- `runHierarchy` — Formal authority levels with escalation. Low-confidence outputs escalate to higher-authority agents.
+
+**Session**
+- `SessionManager` — Named persistent conversation sessions. `swarm.runInSession()` prepends prior context automatically.
+
+**Workflow**
+- `EventFlow` — Event-driven workflow runtime. Steps subscribe to events, emit new ones; execution is queue-driven rather than DAG-fixed.
+
+**Memory**
+- `EpisodicMemory` — Stores specific past interactions with temporal ordering and tag-based recall.
+- `ProceduralMemory` — Stores "how to" procedures with success rate tracking.
+
+**A2A Protocol — v1.0**
+- `kind: 'task'` on `A2ATask`, `ContextId` type alias for cross-task threading.
+- `AgentCard.offline?`, `A2ATaskState` gains `'streaming'`, `A2AMessage` gains `messageId` and `contextId`.
+- `tasks/sendSubscribe` JSON-RPC method for SSE push.
+- `streamSubscribe()` client function.
+- Default `protocolVersion` bumped to `'1.0'`.
+
+**Catalog**
+- `AgentCatalog` — Runtime agent discovery by capability, tag, availability, or semantic description. Heartbeat-based liveness.
+
+**Voice**
+- `VoicePipeline` — STT → LLM → TTS pipeline. Factory methods for Deepgram, ElevenLabs, OpenAI STT/TTS.
+
+---
+
+## [1.3.0] — prior
+
+### New Features
+- Hooks system (`HookRegistry`, priority-ordered hooks, swarm event bridging)
+- Consensus protocols (`RaftNode`, `ByzantineNode`, `GossipNode`)
+- Hive-Mind pattern (`runHiveMind`)
+- Federation hub (`FederationHub`)
+- `ReasoningBank` — trajectory-based pattern memory with EWC
+- Vector quantization (`createQuantizer` — binary, scalar, product)
+- `AttentionRouter` — multi-head attention-based agent routing
+- `RLRouter` / `RLRouterPPO` — reinforcement learning routers
+
+---
+
+## [1.2.0] — prior
+
+### New Features
+- A/B testing engine
+- Judge agent for quality evaluation
+- Weight table for dynamic routing
+- Distillation collector for training pairs (LLMRouter)
+
+---
+
+## [1.1.0] — prior
+
+### New Features
+- Self-learning memory with Elastic Weight Consolidation (EWC)
+- Vector memory with HNSW-like approximate nearest neighbor search
+- 3-tier intelligent model routing
+- Token optimizer (pattern caching, prompt compression, batch optimization)
+- Knowledge graph with PageRank-based importance
+- Background worker system (memory optimizer, pattern learner, metrics, health check)
+- Threat detection system (SQL/command/XSS injection, path traversal, secrets, PII)
+- ADR framework for spec-driven development
+- 10 new agent templates (17 total)
+
+---
+
+## [1.0.0] — initial release
+
+- Budget-first multi-agent orchestration
+- Orchestrator-worker, pipeline, map-reduce, debate, blackboard, fan-out patterns
+- Anthropic, OpenAI, Gemini, Ollama, generic OpenAI-compatible providers
+- Circuit breaker, failover, rate limiter
+- Routing stack: SemanticCache, LatencyRouter, CascadeRouter, SpeculativeCascade, QueryDecomposer
+- MCP tool loading
+- A2A v0.3 protocol
+- Record/Replay testing, evals framework
+- Guardrails (PII, injection, hallucination, length, content filter)
+- Output contracts (schema + semantic validation)
+- Approval gates
+- YAML workflow compiler
+- Dry-run cost projection
+- Differential execution (skip unchanged steps)
+- SSE streaming transport
+- OpenTelemetry export
+- Plugin system
+- 7 agent templates
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -24,29 +24,36 @@ npm run clean                # Remove dist/
 | Module | Path | Purpose |
 |--------|------|---------|
 | **Types** | `src/types/` | All TypeScript interfaces — agent, budget, plan, task, execution, provider, tool, memory, pattern |
-| **Core** | `src/core/` | `Swarm` class, `createAgent()`, MCP tool loader, `MessageBoard` (inter-agent messaging), `stub-board` (no-op board for patterns), guardrails (input/output/tool guards with built-ins: PII, injection, hallucination, max-length, content-filter), output contracts (schema + semantic validation) |
+| **Core** | `src/core/` | `Swarm` class, `createAgent()`, MCP tool loader, `MessageBoard`, `ReputationBoard` (reputation-weighted board), `stub-board`, guardrails, output contracts |
 | **Budget** | `src/budget/` | `BudgetLedger` (hard enforcement), cost optimizer |
-| **Planner** | `src/planner/` | Task scorer, DAG builder, model router, adaptive router, cascade router, semantic cache, speculative cascade, query decomposer, latency router, **3-tier intelligent routing** |
-| **Executor** | `src/executor/` | Parallel DAG runner, checkpoint/resume, dry-run cost projection, differential execution |
-| **Patterns** | `src/patterns/` | Orchestrator-worker, pipeline, map-reduce, debate, blackboard |
-| **Providers** | `src/providers/` | Anthropic, OpenAI, Gemini, Ollama adapters, generic OpenAI-compatible (LiteLLM/vLLM), circuit breaker, failover, rate limiter, model cascade on quality |
+| **Planner** | `src/planner/` | Task scorer, DAG builder, model router, adaptive router, cascade router, semantic cache, speculative cascade, query decomposer, latency router, attention router, RL router, 3-tier routing |
+| **Executor** | `src/executor/` | Parallel DAG runner, checkpoint/resume, dry-run, differential execution, time-travel debugging, rollback manager, trajectory reducer (AgentDiet), speculative tool executor (PASTE) |
+| **Patterns** | `src/patterns/` | Orchestrator-worker, pipeline, map-reduce, debate, blackboard, fan-out, hive-mind, hierarchy, loop-agent |
+| **Providers** | `src/providers/` | Anthropic, OpenAI, Gemini, Ollama, generic OpenAI-compatible (LiteLLM/vLLM), circuit breaker, failover, rate limiter, model cascade on quality |
 | **Conflict** | `src/conflict/` | Contradiction detector (Jaccard/structural), resolver (vote/evidence/escalate) |
 | **Context** | `src/context/` | Token-budget-aware context packer |
-| **A2A** | `src/a2a/` | Agent2Agent protocol — server, client, agent cards |
+| **A2A** | `src/a2a/` | Agent2Agent protocol v1.0 — server, client, agent cards, contextId, streaming state, `tasks/sendSubscribe` |
 | **Pool** | `src/pool/` | Worker pool with lifecycle, concurrency, warm pooling |
-| **Trace** | `src/trace/` | Human-readable execution reports, DAG visualization |
-| **Workflow** | `src/workflow/` | YAML workflow parser + compiler to executable Plans |
-| **Templates** | `src/templates/` | 17 pre-built agent templates (researcher, code-reviewer, synthesizer, data-analyst, qa-tester, writer, planner, security-auditor, devops-engineer, database-engineer, api-designer, performance-engineer, documentation-specialist, architecture-advisor, debugger, refactoring-specialist, integration-specialist, test-automation-engineer) |
+| **Trace** | `src/trace/` | Human-readable execution reports, DAG visualization, OTel export (`toOTelSpans`, `toOTLPJson`), OTel auto-exporter (OTLP push) |
+| **Workflow** | `src/workflow/` | YAML workflow parser + compiler, event-driven workflows (`EventFlow`), graph state machine (`StateMachine`) |
+| **Templates** | `src/templates/` | 17 pre-built agent templates |
 | **Adapters** | `src/adapters/` | Claude Agent SDK wrapper |
-| **Orchestrator** | `src/orchestrator/` | Evolving orchestrator (bandit-based adaptive sequencing), **A/B testing engine**, **Judge agent for quality evaluation**, **Weight table for dynamic routing**, **Distillation collector for training pairs** |
+| **Orchestrator** | `src/orchestrator/` | Evolving orchestrator, A/B testing, judge agent, weight table, distillation collector |
 | **Persistence** | `src/persistence/` | Save/load state to disk or memory backend |
-| **Memory** | `src/memory/` | ANCS memory backend, **self-learning memory with EWC**, **vector memory with HNSW-like search** |
-| **Testing** | `src/testing/` | `RecordingProvider` (wraps real provider, saves fixtures), `ReplayProvider` (loads fixtures, zero-cost deterministic replay), evals framework |
-| **Optimizer** | `src/optimizer/` | Token optimizer with pattern caching, compression, and batch optimization |
-| **Workers** | `src/workers/` | Background worker system for continuous optimization (memory optimizer, pattern learner, metrics collector, cache cleanup, health check) |
-| **Security** | `src/security/` | Threat detection system (SQL/command/XSS injection, path traversal, hardcoded secrets, prompt injection, PII detection) |
-| **Spec** | `src/spec/` | Architecture Decision Records (ADRs) framework for spec-driven development |
-| **Graph** | `src/graph/` | Knowledge graph with PageRank-based importance, graph-enhanced ranked retrieval |
+| **Memory** | `src/memory/` | ANCS, self-learning (EWC), vector (HNSW-like), A-MEM (Zettelkasten graph), temporal (CMA decay), self-editing blocks (Letta), episodic, procedural, external vector store adapters (Pinecone/Qdrant/Redis/flat) |
+| **Session** | `src/session/` | Named persistent sessions, `SessionManager`, conversation branching (`BranchManager`) |
+| **Testing** | `src/testing/` | `RecordingProvider`, `ReplayProvider`, evals framework, `EvalHarness` (run history + regression), trajectory evaluation (TRACE) |
+| **Optimizer** | `src/optimizer/` | Token optimizer, prompt optimizer (DSPy-style) |
+| **Workers** | `src/workers/` | Background workers (memory optimizer, pattern learner, metrics, health check), sleep-time compute agent |
+| **Security** | `src/security/` | Threat detection (SQL/command/XSS injection, path traversal, secrets, prompt injection, PII) |
+| **Tools** | `src/tools/` | Code execution sandbox (Node vm / Docker / E2B), browser tool (Playwright), computer use (Anthropic), skill reducer (progressive disclosure) |
+| **Voice** | `src/voice/` | Voice agent pipeline (STT → LLM → TTS), Deepgram/ElevenLabs/OpenAI providers |
+| **Catalog** | `src/catalog/` | Agent discovery catalog with semantic search |
+| **Hooks** | `src/hooks/` | Hook registry, priority-ordered hooks, swarm event bridging |
+| **Consensus** | `src/consensus/` | Raft, Byzantine fault-tolerant, Gossip consensus |
+| **Federation** | `src/federation/` | Multi-swarm federation hub |
+| **Spec** | `src/spec/` | Architecture Decision Records (ADRs) |
+| **Graph** | `src/graph/` | Knowledge graph with PageRank, graph-enhanced retrieval |
 
 ### Data Flow
 
@@ -130,7 +137,7 @@ Task → Scorer → Planner (DAG) → Executor
 
 - All tests in `tests/unit/` — pure unit tests, no external services
 - Tests use mock providers that return canned responses
-- 29 test files, 265 tests
+- 71 test files, 621 tests
 - Run with `npm test`
 
 ## Peer Dependencies