v0.2.6

[0.2.6] - 2026-07-07

New Features

• Unified SDK Package (sdk/): New top-level API sdk.MustNew() / sdk.New() with functional options (WithOpenAI, WithOllama, WithAnthropic, WithDefaultMemory, WithEvolution, WithMCP, WithHumanInput, etc.). Single entry point for LLM, tools, memory, evolution, and MCP.
• Agent Runtime: agent.Run(ctx, input) ReAct loop with tool calling, memory context injection, token tracking, and result metadata.
• Streaming Support: agent.Stream(ctx, input) returns <-chan StreamChunk for async response streaming.
• Multi-Agent Teams: rt.NewTeam(name, leader, members) with team.Run() for leader/member orchestration.
• Human-in-the-Loop: WithHumanInput() callback for tool call approval before execution.
• MCP Integration: WithMCP() connects to MCP servers via stdio, auto-registers their tools.
• Strategy Evolution: rt.Evolve(ctx, agent, task) evolves agent instructions via LLM. WithEvolution() enables the evolution system.
• CLI Tools (cmd/ares/): ares init (scaffold project), ares run (run agent from config, auto-detects ares.yaml), ares bench (benchmark with JSON/Markdown output), ares doctor (diagnose environment), ares version.
• Config-Driven Setup: sdk.LoadConfigFile(path) reads YAML config, cfg.ToOptions() converts to SDK options. ares run auto-discovers ares.yaml or config/ares.yaml.
• Evaluation Framework (evaluation/): evaluation.New(), Register(), RunScenario(), RunAll() with structured Metrics, Report, Aggregate. Report output via ToMarkdown() / ToJSON(). Built-in scenarios: basic-chat, tool-calling, multi-agent, resilience, evolution.

Examples

• 9 New SDK Examples: Numbered 01-quickstart through 09-full-app, each with ares.yaml config.
  • 01-quickstart: Minimal agent in 20 lines
  • 02-tool-calling: Multi-tool registration
  • 03-dag-workflow: MutableDAG + conditional branching
  • 04-multi-agent: Leader/member team orchestration
  • 05-evolution-demo: Instruction evolution before/after comparison
  • 06-chaos-resilience: 9 failure modes (file, timeout, network, MCP, LLM, memory, graceful degradation)
  • 07-human-in-loop: Tool call approval with WithHumanInput
  • 08-mcp-integration: MCP server connection via WithMCP
  • 09-full-app: Web UI + Agent + Tools + Memory + Stats dashboard
• Evaluation Example (examples/eval/): Runs all 5 capability scenarios with scoring.

Documentation

• README Rewrite: Reduced from 774 to 214 lines. SDK Quick Start at the top. English (README.md) and Chinese (README_CN.md) versions.
• GitHub Pages Website: docs/index.html with dark theme, marked.js inline Markdown rendering, all articles browsable.
• Architecture Diagram: Mermaid diagram covering SDK, LLM providers, Tools, Memory, Evolution, CLI, Examples.
• 7 Cookbook Recipes: docs/cookbook/ with Chat, Tool Calling, Multi-Agent, Memory, Coding Agent, Code Review, GitHub Agent.
• CI Docs Deployment: GitHub Actions workflow (docs.yml) auto-deploys docs/ to Pages.

Code Quality

• SDK Test Coverage: 54%+ with 20+ tests covering Runtime, Agent, Team, Config, Evolution, Streaming, MCP, HumanInput, Benchmarks. All pass with go test -short ./....
• Lint Clean: golangci-lint 0 issues across SDK, CLI, evaluation, and examples.
• English Comments: All code comments in English per code_rules.md.
• Binary Rename: CLI binary ARES → ares (lowercase).

Infrastructure

• Docker Compose: docker-compose.yml + Dockerfile.demo for one-command demo deployment (Ollama + full-app).
• Makefile: Added quickstart, examples, install-cli, test-eval targets.
• Example Cleanup: Removed 20+ stale/duplicate examples; kept 9 curated SDK examples + advanced ones in git history.
• Chaos Arena YAML: Restored examples/arena/leader_assassination.yaml and cascading_storm.yaml with all built-in action types.

Performance

• GA Diversity Sampling: Added DiversitySampleSize config (default 200) to estimate numeric diversity via random neighbor sampling instead of O(n²) exact computation. Stats(pop=1000) latency dropped 38% (69.5ms → 43.3ms). Configurable per PopulationConfig.DiversitySampleSize.
• Fitness Sharing Optimization: Replaced per-agent Fisher-Yates full permutation with Reservoir Sampling in applyFitnessSharingSampled. Allocation reduced 44% for all population sizes (pop=100: 185→106 allocs, pop=500: 905→506 allocs). GC pressure halved in large evolution runs.
• Subscribe Allocation Reduction: Replaced UUID subscription IDs with atomic.Int64 counter and removed *sync.Once per subscriber. Allocs reduced 33% (900→600 per 100 subscribers). Channel buffer increased from 1→64 to reduce burst drops.
• Benchmark Report: Comprehensive benchmark report across all modules (events, GA genome/evaluation, memory distillation, tools core, handlers, errors) with full platform config (Apple M3 Max, Go 1.26, 3-run average).

New Features

• Memory Pipeline Complete: End-to-end memory pipeline with ReportGenerator, PushService, and report formatting for human-readable evolution summaries. Full cycle: evaluation → distillation → report → push.
• Agent Age Eviction: AgentMaxAge config limits strategy lifespan; GenerationCreated tracking ensures agents survive exactly AgentMaxAge generations. Legacy strategies (GenerationCreated==0) exempted.
• Confidence Calculation: Added sample-based confidence to AggregateEvidenceCrossTask, enabling evidence quality scoring in cross-task aggregation.

Refactors

• Truncate Utility Consolidation: Unified internal/ares_memory/internal/truncate package for reusable truncation logic across memory and LLM modules.
• Evidence Logic Cleanup: AggregateEvidence refactored for clarity; cross-task evidence aggregation now filters mixed-task noise with AggregateEvidenceCrossTask.
• FIXME Cleanup (22 files): Removed stale FIXME comments in internal/ares_quant/, internal/api_impl/, api/client/, internal/ares_events/, internal/storage/postgres/services/. All had already been implemented but comments were not updated.
• Promotion Logic: Tightened statistical bands (5-20x → 6-18x) in selection_extra_test.go and reduced low-scorer threshold (5% → 0.2%) for more deterministic selection verification.

Bug Fixes

• Ignored json.Marshal Errors: Fixed 4 ignored json.Marshal calls in internal/ares_events/summary_repository.go — previously would silently produce null DB values on serialization failure. Now errors propagate with fmt.Errorf("marshal %s: %w", ...).
• Errgroup Context Propagation: In internal/api_impl/service.go, errgroup.WithContext(ctx) returned a derived context cancelled on sibling errors — but it was discarded with _. Fix: s.g, s.ctx = errgroup.WithContext(ctx) to enable proper error propagation.
• SSE Health Probe: Implemented real SSE health check via ConnectSSE instead of hardcoded assumed healthy.
• Generation Logging: Fixed generation=0 in logs by using absolute Population.Generation in callback_gen.
• GenerationCreated Off-by-One: Use Generation+1 so agents survive exactly AgentMaxAge generations.
• Guardrail Config Default: Inverted PromptDiversityGuardEnabled → DisablePromptDiversityGuard (default enabled).

v0.2.5

[0.2.5] - 2026-07-02

Performance

• GA Diversity Sampling: Added DiversitySampleSize config (default 200) to estimate numeric diversity via random neighbor sampling instead of O(n²) exact computation. Stats(pop=1000) latency dropped 38% (69.5ms → 43.3ms). Configurable per PopulationConfig.DiversitySampleSize.
• Fitness Sharing Optimization: Replaced per-agent Fisher-Yates full permutation with Reservoir Sampling in applyFitnessSharingSampled. Allocation reduced 44% for all population sizes (pop=100: 185→106 allocs, pop=500: 905→506 allocs). GC pressure halved in large evolution runs.
• Subscribe Allocation Reduction: Replaced UUID subscription IDs with atomic.Int64 counter and removed *sync.Once per subscriber. Allocs reduced 33% (900→600 per 100 subscribers). Channel buffer increased from 1→64 to reduce burst drops.
• Benchmark Report: Comprehensive benchmark report across all modules (events, GA genome/evaluation, memory distillation, tools core, handlers, errors) with full platform config (Apple M3 Max, Go 1.26, 3-run average).

New Features

• Memory Pipeline Complete: End-to-end memory pipeline with ReportGenerator, PushService, and report formatting for human-readable evolution summaries. Full cycle: evaluation → distillation → report → push.
• Agent Age Eviction: AgentMaxAge config limits strategy lifespan; GenerationCreated tracking ensures agents survive exactly AgentMaxAge generations. Legacy strategies (GenerationCreated==0) exempted.
• Confidence Calculation: Added sample-based confidence to AggregateEvidenceCrossTask, enabling evidence quality scoring in cross-task aggregation.

Refactors

• Truncate Utility Consolidation: Unified internal/ares_memory/internal/truncate package for reusable truncation logic across memory and LLM modules.
• Evidence Logic Cleanup: AggregateEvidence refactored for clarity; cross-task evidence aggregation now filters mixed-task noise with AggregateEvidenceCrossTask.
• FIXME Cleanup (22 files): Removed stale FIXME comments in internal/ares_quant/, internal/api_impl/, api/client/, internal/ares_events/, internal/storage/postgres/services/. All had already been implemented but comments were not updated.
• Promotion Logic: Tightened statistical bands (5-20x → 6-18x) in selection_extra_test.go and reduced low-scorer threshold (5% → 0.2%) for more deterministic selection verification.

Bug Fixes

• Ignored json.Marshal Errors: Fixed 4 ignored json.Marshal calls in internal/ares_events/summary_repository.go — previously would silently produce null DB values on serialization failure. Now errors propagate with fmt.Errorf("marshal %s: %w", ...).
• Errgroup Context Propagation: In internal/api_impl/service.go, errgroup.WithContext(ctx) returned a derived context cancelled on sibling errors — but it was discarded with _. Fix: s.g, s.ctx = errgroup.WithContext(ctx) to enable proper error propagation.
• SSE Health Probe: Implemented real SSE health check via ConnectSSE instead of hardcoded assumed healthy.
• Generation Logging: Fixed generation=0 in logs by using absolute Population.Generation in callback_gen.
• GenerationCreated Off-by-One: Use Generation+1 so agents survive exactly AgentMaxAge generations.
• Guardrail Config Default: Inverted PromptDiversityGuardEnabled → DisablePromptDiversityGuard (default enabled).

Code Quality

• Unit Test Coverage (service.go): Added internal/ares_evolution/service/service_test.go (383 lines, 23 test cases). Coverage of service.go increased 16.3% → 47.2%. Key functions: NewService 93.3%, Evolve 80.4%, toAPIStrategy/toInternalStrategy/cloneDimensionScores all 100%.
• LLM Scorer Tests: Added pure-logic tests for extractScoreFromText, fallbackScore, buildPrompt, parseScore (35 table-driven test cases). No LLM required.
• Task Planner Tests: Consolidated 10 repetitive test functions into 2 table-driven tests with meaningful result.Error content assertions. TestFormatToolsList converted to table-driven.
• Test Weakness Assessment: Sampled 20+ non-testify test files — confirmed they have meaningful multi-field assertions (not perfunctory). Postgres integration tests properly isolated behind //go:build integration.
• Docker Compose Update: Added Ollama service for local LLM fallback in development stack. Updated benchmark links in both EN and CN README.

What's Changed

• 0.2.5 by @Timwood0x10 in #36

Full Changelog: v0.2.4...v0.2.5

v0.2.4

[0.2.4] - 2026-06-28

New Features

• Plugin System Architecture: Full plugin system with PluginBus, RuntimePlugin interface, WorkflowHook interface, and capability-based plugin discovery. 10 built-in plugins: ObserverPlugin, CheckpointPlugin, ToolPlugin, ExpressionRouter, MemoryRouter, EvolutionRouter, LoopPlugin, RecoveryPlugin, InterruptPlugin, ArenaPlugin.
• Genetic Algorithm Evolution System (Beta): Complete GA package with Population, Crossover (Inherit/HalfSplit/Uniform), TournamentSelection, strategy mutation engine, diversity tracking with fitness sharing, adaptive survival rates, and deterministic reproduction via seed control.
• Autonomous Evolution (Dream Mode v1): Closed-loop evolution orchestration with Dream Cycle (trigger → mutate → evaluate → adopt → record lineage). Arena regression testing with Welch's t-test, bandit feedback loop for experience quality optimization, and full genealogy tracking.
• Batch LLM Scorer: Concurrent LLM scoring with failover resilience for evolution pipeline.
• ExecuteFromCheckpoint: Lightweight workflow resume from checkpoint via Graph.ExecuteFromCheckpoint(). Checkpoint integration via PluginBus hooks.
• LoopPlugin: Controlled execution loops with configurable iteration limits.
• RouterPlugin Auto-Wiring: Automatic plugin registration based on declared capabilities.
• Execution Collector: Thread-safe runtime data aggregation for route recording and tool invocation tracking.
• Module-Scoped Structured Logging: Each core module emits logs with module field for traceability. Added logger.Module() helper in internal/logger/. 12 core packages converted.
• Event ModuleName Field: Added ModuleName field to Event struct. Emit() and PluginBus.Emit() now accept moduleName parameter for full traceability of which module emitted each event.
• Abstract API Layer: Added interfaces in api/core/ for all major modules: AgentService, Runtime, WorkflowService, MemoryService, LLMService, RetrievalService, Evolution, DreamCycle, Arena, ContextCleaner.
• Bootstrap Factory: Added api/bootstrap/ package that wires all ARES modules (Runtime, Memory, Evolution, Arena, EventStore) into a single ARES container with New(), Start(), Stop(), RunEvolution(), and ExecuteArenaAction().
• Interview Demo: Complete interview demo stack with web search tool and prompt length validation.
• JSONL Training Data Pipeline: End-to-end pipeline for agent strategy evolution and experience distillation data export.

Refactors

• Unified Package Naming: Renamed 15 internal packages to ares_ prefix: bootstrap, callbacks, ctxutil, shutdown, ratelimit, security, config, eval, observability, integration, events, mcp, protocol, quant, runtime.
• API Layer Thinning: Moved all independent service implementations from api/ to internal/. The api/ layer now only contains interface definitions, error types, HTTP handlers, router, and client SDK. Moved packages:
  • api/service/agent → internal/agents/
  • api/service/graph → internal/workflow/graphservice/
  • api/service/llm → internal/llmservice/
  • api/service/memory → internal/memoryservice/
  • api/service/retrieval → internal/retrievalservice/
  • api/ares_evolution → internal/ares_evolution/service/
  • api/ares_memory → internal/ares_memory/service/
  • api/ares_retrieval → internal/ares_memory/retrieval_api/
  • api/ares_experience → internal/ares_experience/service/
  • api/eval → internal/ares_eval/service/
  • api/marketmaking → internal/ares_quant/marketmaking_api/
• Evolution Genome Wiring: Split genome_wiring into separate module, fix guardrails, wire dream cycle.
• HITL Feedback Plugin: Moved from standalone to workflow engine integration.
• Graph Builder APIs: Migrated all graph builder APIs to return errors instead of panicking.
• Scoring Cache: Replaced sync.RWMutex with atomic counters for hit/miss tracking.
• Evolution Mutation: Restructured mutation logic with experience-guided evolution system.
• Performance: Increased concurrent LLM scoring limits and optimized sampling. Completed P0/P1/P2 performance improvements.

Bug Fixes

• Replaced time.Sleep with channel-based event test pattern in graph executor tests.
• Fixed indentation in executor_test.go.
• Fixed data race in DynamicExecutor recovery path with proper timeout handling.
• Fixed MemoryEventStore.Close() idempotency — second+ calls return ErrEventStoreClosed.
• Fixed SSE transport double resp.Body.Close() causing panic on shutdown.
• Fixed LLM client Close() race condition via sync.Once.
• Fixed OpenAI adapter silently swallowing io.ReadAll errors in error paths.
• Fixed Population.ScoreAgents panic recovery logging with agent context.
• Fixed updateBestEverLocked concurrency safety with deep copy via a.Clone().
• Fixed NewTaskPlanner/NewTaskPlannerWithConfig silent fallback from invalid maxTasks.
• Added nil validation in leader.New, NewTaskDispatcher, and NewMCPManager.

What's Changed

• 0.2.4 by @Timwood0x10 in #35

Full Changelog: v0.2.3...v0.2.4

v0.2.3

[0.2.3] - 2026-06-24

New Features

• Genetic Algorithm Evolution System (Beta): Full GA genome package with Population, Crossover (Inherit/HalfSplit/Uniform modes), TournamentSelection, and strategy mutation engine. Supports deterministic reproduction via seed control, elite preservation, adaptive survival rates, and diversity tracking with fitness sharing. (GA Hardening Plan)
• Autonomous Evolution (Dream Mode v1): Closed-loop evolution orchestration with Dream Cycle (trigger → mutate → evaluate → adopt → record lineage). Includes arena regression testing with Welch's t-test, bandit feedback loop for experience quality optimization, and full genealogy tracking.
• Agent Resurrection & Snapshot System: Pluggable health checking for agent recovery, checkpoint-based resurrection with state restoration from EventStore and MemoryStore.
• Tiered Scoring System: Multi-level scoring pipeline with FailoverScorer integration. Includes scoring cache optimization (atomic hit/miss counters), hybrid scoring with prompt crossover modes, and unevaluated score guardrails.
• JSONL Training Data Pipeline: End-to-end pipeline for agent strategy evolution and experience distillation data export.
• Leader Agent Hardening: Nil validation for all constructor parameters (memory manager, aggregator, parser, planner, dispatcher). Session initialization via sync.Once. Comprehensive error collection during Stop() with joined errors from distillation/streaming goroutines.
• Workflow Engine Hardening: Thread-safe HITL handler/store access via sync.RWMutex. Workflow execution timeout (default 30s) to prevent indefinite blocking. Proper OutputStore.Close() cleanup.
• Event Store Hardening: Errgroup-based compaction with timeout context (30s). Nil compactor/repo guards in all read paths. MemoryEventStore Close() returns ErrEventStoreClosed on double-close for idempotent shutdown.
• LLM Client Validation: Config validation enforces required Provider and BaseURL fields. Close() idempotency via sync.Once. OpenAI adapter properly handles io.ReadAll errors instead of silently discarding them.
• MCP Client Hardening: Nil client guard in tool registration. Godoc-style documentation for all public APIs. SSE transport fixes double-close of resp.Body (deferred close only in receiveLoop).
• Memory Manager Config Validation: Validates MaxTasks, MaxDistilledTasks, DistilledTaskTTL, and VectorDim are positive. Stop() collects all errors and returns them joined.
• Crossover & Selection Validation: Crossover.Validate() and TournamentSelection.Validate() methods for post-construction config invariance checking. Defensive nil checks and enum validation.

Improvements

• Renamed project to ARES (Adaptive Resilient Evolution System)
• Enhanced scoring cache with atomic counters replacing sync.RWMutex
• Improved error visibility with structured error wrapping across all modules
• Added debug logging to TaskDispatcher and planner fallback warnings
• Default DistilledTaskTTL set to 30 days in DefaultMemoryConfig()
• Guarded all CompactableEventStore read paths against nil compactor/repo

Bug Fixes

• Fixed data race in DynamicExecutor recovery path with proper timeout handling
• Fixed MemoryEventStore.Close() idempotency — second+ calls return ErrEventStoreClosed
• Fixed SSE transport double resp.Body.Close() causing panic on shutdown
• Fixed LLM client Close() race condition via sync.Once
• Fixed OpenAI adapter silently swallowing io.ReadAll errors in error paths
• Fixed Population.ScoreAgents panic recovery logging with agent context
• Fixed updateBestEverLocked concurrency safety with deep copy via a.Clone()
• Fixed NewTaskPlanner/NewTaskPlannerWithConfig silent fallback from invalid maxTasks
• Added nil validation in leader.New, NewTaskDispatcher, and NewMCPManager
• fix #33
• fix #32

What's Changed

• 0.2.3 by @Timwood0x10 in #34

Full Changelog: v0.2.2...v0.2.3

v0.2.2

[0.2.2] - 2026-06-19

New Features

• Embedding Lifecycle Unification: Unified embedding lifecycle across distillation, storage, and retrieval pipelines. Embedding workflows now share a common lifecycle model, reducing code duplication and ensuring consistent behavior during creation, update, and deletion of embeddings.
• Context Cleaning: Automatic context window management with tool call causality preservation. The context cleaner maintains causal ordering of tool calls during cleanup, preventing out-of-order execution after context truncation.
• Workflow Enhancements: MutableDAG.ReplaceNode for replacing nodes at runtime. Custom RecoveryHandler for failure recovery per workflow step. Enhanced event propagation across workflow execution.
• Portfolio Simulator: Investment portfolio simulation system with multi-asset backtesting. Includes research memory bridge connecting portfolio simulation results to the research memory system for data-driven investment decisions.
• Investment Simulator: Standalone investment simulation module for modeling and analyzing investment strategies.
• CoinGecko Crypto Feed: Real-time cryptocurrency price data integration via CoinGecko API, enabling live market data for trading analysis.
• Public Marketmaking API: Marketmaking API migrated from internal to public (api/marketmaking/), with multi-asset backtesting support. Includes comprehensive paper trading, chaos testing, and backtesting capabilities.
• Quant Trading Example: Complete quantitative trading example with SQLite backend, demonstrating end-to-end quant trading workflow.
• Tool Lifecycle Events: Emit lifecycle events for tool execution, enabling observability and monitoring of tool calls throughout their lifecycle.
• Memory Metadata Propagation: Expanded metadata propagation across memory operations, enriching context with session and agent metadata.
• Concurrent Distillation Pipeline: errgroup-based parallel embedding in distiller (concurrency limit 5), and concurrent experience storage in manager_impl.go, reducing end-to-end distillation latency.
• Content Hash Dedup: SHA-256 content_hash column on distilled_memeries with ON CONFLICT (tenant_id, content_hash) WHERE content_hash IS NOT NULL DO NOTHING for idempotent memory storage.
• Idempotent Migrations: All DDL operations now safe to re-run — DROP IF EXISTS + CREATE for policies/triggers, IF NOT EXISTS for indexes, ADD COLUMN IF NOT EXISTS for schema evolution.
• Chinese Language Support: Chinese keyword detection in detector.go (介绍/是什么/怎么/有哪些/区别/说说/推荐 etc.) and Chinese importance scoring in scorer.go (错误/修复/配置/框架/架构/优化 etc.), enabling experience extraction from Chinese Q&A.
• Knowledge Correction Flow: End-to-end correction pipeline in knowledge-base example — detects correction intent, calls LLM for structured commands (UPDATE:/DELETE:/CREATE:), executes DB writes for both distilled_memories and knowledge_chunks_1024. Supports correction via "纠错" keyword.
• RAG Search Includes Corrected Memories: KnowledgeBase.Search() now queries both knowledge_chunks_1024 and distilled_memories, with corrected memories boosted in ranking.
• Restart Script Import: scripts/docker/restart.sh --save <path> option to import a document immediately after DB migration.

Refactors

• Enhanced configuration safety with improved validation and error handling in the API layer (api/config.go, api/service.go).
• Renamed quant-demo to quant-trading with updated configuration and documentation.
• Enforced snapshot-only data constraint in analyst prompts to ensure data consistency.
• Replaced WriteString(fmt.Sprintf(...)) with fmt.Fprintf across correction flow.
• Simplified loop with append(..., distilledResults...).

Bug Fixes

• Resolved data race and timing issues in DynamicExecutor recovery path.
• Fixed documentation file naming inconsistencies.
• Fixed all errcheck issues (unchecked Close() calls) in cmd/ migration tools.
• Fixed De Morgan's law simplification in UUID validation.

Acknowledgements

We sincerely thank the iflow.cn community members for their constructive feedback on the project.

• @chigefeijimu made big idea
  @Timwood0x10

v0.2.1

New Features

• MCP Client: Model Context Protocol client implementation with JSON-RPC 2.0 messaging, stdio and SSE transport support, tool schema management, and connection lifecycle management.
• Web Dashboard: Real-time monitoring dashboard with WebSocket hub, REST API v2, orchestrator for multi-agent coordination, event bridge for system state streaming, and static asset serving.
• Flight Recorder: Multi-agent runtime intelligence recording with timeline tracking, decision logging, diagnostics engine, agent genealogy graph, DOT/JSON export, and replay pipeline.
• Chaos Engineering Arena: Fault injection framework with injector supporting process_kill, network_partition, latency_spike, and kill_orchestrator fault types; resilience scoring with configurable metrics; survival mode for continuous chaos testing; HTTP API and YAML scenario configuration.
• Callbacks System: Event-driven callback mechanism with typed event contexts, handler registry, and lifecycle hooks for agent/tool/runtime events.
• LLM Output Parsing: Multi-provider output adapters (OpenAI, Ollama, OpenRouter), prompt template engine with Go template syntax, function calling extraction and validation, schema-based parameter validation, and streaming output parser.
• Function Calling: LLM function calling support with tool schema generation, argument extraction, and result formatting.
• Agent Genealogy: Agent lineage tracking with parent-child relationships, birth/death event recording, and genealogy graph export.
• Event Auto-Compaction: Configurable event store compaction with retention policies, snapshot-based trimming, and automatic execution.
• Tool Lifecycle Hooks: Pre/post execution hooks for tools with context injection and error handling.
• Quant Demo: Quantitative analysis example with CSV data processing.
• DevAgent Example: Development agent example with workflow configuration.
• MCP Dashboard Example: Dashboard integration example with MCP transport.
• Capability Demo: Tool capability demonstration example.

Improvements

• Pruned unused components and deduplicated code across runtime resurrection module
• Improved error visibility with structured error messages
• Extracted restore logic into reusable functions
• Exposed migration DDL for external tooling
• Cleaned up validators and reduced code duplication
• Streamlined dashboard frontend assets
• Generalized domain models for broader use cases beyond original fashion domain

Bug Fixes

• Fixed various lint issues identified by golangci-lint
• Added GetAgent method to Runtime interface
• Wired verifyRestoredState in example code
• Corrected semaphore available count calculation
• Escaped password in DSN connections
• Added ILIKE pattern escaping for PostgreSQL queries

v2.0.0 (2026-06-11)

New Features

• Leader Failover: Checkpoint-based recovery with LeaderSupervisor detecting leader failure, recovering stale tasks from last checkpoint, and reassigning work to available sub-agents. ColdRestartStrategy for deterministic recovery.
• Runtime Dynamic Graph: MutableDAG with thread-safe mutation (add/remove nodes and edges at runtime). DynamicExecutor with ApplyMode for hot-reload without stopping execution. Incremental cycle detection on edge insertion.
• Human-in-the-Loop: InterruptConfig on workflow steps for human approval gates. InterruptHandler blocks execution until approved. InterruptStore provides crash recovery of pending approvals.
• Agent Resurrection Plugin: Pluggable HealthChecker interface for custom health detection. HeartbeatAdapter for heartbeat-based liveness. Supervisor for automatic agent restart on failure.
• Event Sourcing: EventStore interface with optimistic concurrency control. MemoryEventStore for dev/test, PostgresEventStore for production. 17 event types covering agent lifecycle, tasks, sessions, workflows, and failover. Pub/sub via Subscribe with filtered event channels. DLQ auto-retry with configurable retry budgets.
• Pluggable Vector Store: VectorStore interface replacing concrete *VectorSearcher in Repository. PostgreSQL + pgvector for production, in-memory for dev/test. Drop-in replacement support for Qdrant, Milvus, SQLite, or custom backends.
• WorkflowService API: High-level workflow orchestration abstraction over the DAG engine.

Bug Fixes (46 fixes)

Storage (12 fixes)

• C1: Embedding queue dedup key mismatch causing duplicate embeddings
• C2: Write buffer data loss on Stop() before flush completes
• C3: Embedding enqueue outside transaction leading to orphaned records
• C4: FetchPendingTasks lock ineffective with FOR UPDATE SKIP LOCKED
• C5: Reconcile threshold time arithmetic off by orders of magnitude
• M1: ManagedRow connection leak on error paths
• M2: Missing migration tables in migrate.go
• M3: Circuit breaker halfOpenInflight counter leak
• M4: VectorSearcher missing dimension validation
• M6: FileWatcher TOCTOU race in scanAndLoad
• M7: Map reference shared unsafely in callbacks
• M8: Graph Edge() no validation of node endpoints

Workflow (8 fixes)

• C6: Panic recovery ordering in executor.go (recovery after cleanup)
• C7: Graph executor in-degree tracking incorrect after node removal
• H1: Deadlock false positive in executor.go (errgroup misuse)
• H2: DynamicExecutor hang on node removal during execution
• H3: stepEg.Wait() concurrent with Go() causing race
• H4: NewDAG silently dropping duplicate step IDs
• M5: MaxAttempts=0 skips execution entirely
• M9: recomputeOrder version-check race on concurrent access

AHP Protocol (7 fixes)

• C8: Queue send on closed channel panic during shutdown
• C9: HeartbeatSender Start/Stop race condition
• H5: getRandomSuffix nil dereference on empty slice
• H6: SendMessage swallows all errors silently
• H7: Protocol has no Close() method (resource leak)
• M10: Peek() non-atomic read (race under concurrent access)
• M12: DLQ.Remove leaks trailing pointer after deletion

What's Changed

• v0.2.1 by @Timwood0x10 in #30

Full Changelog: v0.2.0...v0.2.1

v0.2.0

v0.2.0

New Features

• Leader Failover: Checkpoint-based recovery with LeaderSupervisor detecting leader failure, recovering stale tasks from last checkpoint, and reassigning work to available sub-agents. ColdRestartStrategy for deterministic recovery.
• Runtime Dynamic Graph: MutableDAG with thread-safe mutation (add/remove nodes and edges at runtime). DynamicExecutor with ApplyMode for hot-reload without stopping execution. Incremental cycle detection on edge insertion.
• Human-in-the-Loop: InterruptConfig on workflow steps for human approval gates. InterruptHandler blocks execution until approved. InterruptStore provides crash recovery of pending approvals.
• Agent Resurrection Plugin: Pluggable HealthChecker interface for custom health detection. HeartbeatAdapter for heartbeat-based liveness. Supervisor for automatic agent restart on failure.
• Event Sourcing: EventStore interface with optimistic concurrency control. MemoryEventStore for dev/test, PostgresEventStore for production. 17 event types covering agent lifecycle, tasks, sessions, workflows, and failover. Pub/sub via Subscribe with filtered event channels. DLQ auto-retry with configurable retry budgets.
• Pluggable Vector Store: VectorStore interface replacing concrete *VectorSearcher in Repository. PostgreSQL + pgvector for production, in-memory for dev/test. Drop-in replacement support for Qdrant, Milvus, SQLite, or custom backends.
• WorkflowService API: High-level workflow orchestration abstraction over the DAG engine.

Bug Fixes (46 fixes)

Storage (12 fixes)

• C1: Embedding queue dedup key mismatch causing duplicate embeddings
• C2: Write buffer data loss on Stop() before flush completes
• C3: Embedding enqueue outside transaction leading to orphaned records
• C4: FetchPendingTasks lock ineffective with FOR UPDATE SKIP LOCKED
• C5: Reconcile threshold time arithmetic off by orders of magnitude
• M1: ManagedRow connection leak on error paths
• M2: Missing migration tables in migrate.go
• M3: Circuit breaker halfOpenInflight counter leak
• M4: VectorSearcher missing dimension validation
• M6: FileWatcher TOCTOU race in scanAndLoad
• M7: Map reference shared unsafely in callbacks
• M8: Graph Edge() no validation of node endpoints

Workflow (8 fixes)

• C6: Panic recovery ordering in executor.go (recovery after cleanup)
• C7: Graph executor in-degree tracking incorrect after node removal
• H1: Deadlock false positive in executor.go (errgroup misuse)
• H2: DynamicExecutor hang on node removal during execution
• H3: stepEg.Wait() concurrent with Go() causing race
• H4: NewDAG silently dropping duplicate step IDs
• M5: MaxAttempts=0 skips execution entirely
• M9: recomputeOrder version-check race on concurrent access

AHP Protocol (7 fixes)

• C8: Queue send on closed channel panic during shutdown
• C9: HeartbeatSender Start/Stop race condition
• H5: getRandomSuffix nil dereference on empty slice
• H6: SendMessage swallows all errors silently
• H7: Protocol has no Close() method (resource leak)
• M10: Peek() non-atomic read (race under concurrent access)
• M12: DLQ.Remove leaks trailing pointer after deletion

Agent System (8 fixes)

• L6: Start partial validation cleanup (inconsistent state on error)
• L7: SubAgent ProcessStream goroutine leak on context cancellation
• L8: doFailover uses cancelled ctx for Stop (fails to clean up)
• L9: Dispatcher partial results misleading (reports success on partial failure)
• L10: Dynamic executor uses bare go instead of stepEg
• L11: Missing SnapshotWithSteps() on MutableDAG
• M11: NewTaskMessage allows nil payload
• BUG-5: Dead verification in pg_store_test ConcurrentAppend

Event Sourcing (5 fixes)

• BUG-1: FromVersion inclusive/exclusive boundary wrong in memory_store.go
• BUG-2: Since filter inclusive boundary wrong in pg_store.go
• BUG-3: ReadAll incorrectly applies FromVersion filter in memory_store
• BUG-4: Append does not return ErrVersionConflict on unique violation in pg_store
• STYLE-1: Bare go keyword without context cancellation in event store

Other (6 fixes)

• L1: safeFormatTable returning empty string on valid input
• L2: Missing immediate retry after flush failure in write buffer
• F1: workflow_test.go config mismatch with actual types
• Executor nil pointer check on context cancellation (commit 9565f79)
• STYLE-2: Channel buffer size 64 changed to 1 (backpressure)
• STYLE-3: Empty slice literals replaced with nil returns

Infrastructure

• CI/CD pipeline via GitHub Actions (lint, test, race detection, build)
• Integration tests for workflow engine (5 test cases)
• Benchmark suite (32 benchmarks across 8 categories)
• Bilingual documentation (English + Chinese) in docs/en/ and docs/zh/
• Reorganized documentation into language-specific directories
• .golangci.yml configuration for consistent linting

Breaking Changes

• NewLeaderSupervisor signature changed: added eventStore parameter
• NewColdRestartStrategy signature changed: added checkpoint parameter
• MemoryManager interface added GetLatestSessionForLeader method
• VectorStore interface replaces concrete *VectorSearcher in Repository
• NewResultAggregator signature changed: added sortBy string parameter
• TaskPlanner.Plan signature changed: added inputText string parameter
• ResultAggregator.Aggregate signature changed: added tasks []*models.Task parameter
• Domain types renamed: FashionFilters -> ResourceFilters, FashionItem -> ResourceItem, AgentProfile -> AgentUserProfile, AgentRecommendation -> TaskRecommendation, OutfitSuggestion -> Suggestion, AgentTrend -> Trend

What's Changed

• v0.2.0 by @Timwood0x10 in #26

Full Changelog: v0.1.1...v0.2.0

v0.1.0

What's Changed

• Architecture upgrade by @Timwood0x10 in #1
• Improve by @Timwood0x10 in #2
• Feature Enhancement & Bug Fix by @Timwood0x10 in #3
• add SimpleRetrievalService by @Timwood0x10 in #4
• Agent Capability Engine by @Timwood0x10 in #5
• Implement Experience System with bilingual testing support by @Timwood0x10 in #6
• Impl Graph Dynamic Orchestration System by @Timwood0x10 in #7
• FIx BUG by @Timwood0x10 in #8
• Fix bug by @Timwood0x10 in #9
• fix license by @Timwood0x10 in #10
• Fix critical bugs, improve concurrency safety, and complete fashion decoupling by @Timwood0x10 in #19

New Contributors

• @Timwood0x10 made their first contribution in #1

Full Changelog: https://github.com/Timwood0x10/goagent/commits/v0.1.0