Skip to content

openJiuwen agent-core v0.1.18

Latest
Latest

Choose a tag to compare

View all tags
@openjiuwen-release-bot openjiuwen-release-bot released this 11 Sep 08:00
v0.1.18
1d37ae3

openJiuwen agent-core v0.1.18

Release Date: September 9, 2026

Welcome to v0.1.18! This release focuses on runtime governance, team execution reliability, permission architecture, and agent extensibility. Key highlights include SwarmFlow runtime controls and budget operators, team plan mode and ReliabilityRail, a redesigned Permission Engine, atomic Skill capability-graph updates, AgentTemplate and Plugin loading, structured execution traces, and an improved online RL workflow. This release also fixes compatibility and default-parameter handling in the jiuwen-memory provider.

🚀 Features

Agent Teams

  • Agent teams plan mode – Teams can now submit a plan and wait for approval before execution. Plan mode blocks file writes, task creation, and message sending while preserving the existing TUI/Web approval experience (#369).
  • Agent Teams ReliabilityRail – Added repeat-tool, model-error, tool-error, output-length, compaction, and ping-pong detectors with configurable severity actions, leader escalation, local correction, and user notification (#169).
  • Third-party Agent reliability – Added retry and user-guidance handling for transient connection/reading timeouts, rate limiting (429), and service-unavailable (5xx) failures, with unified final error codes (#776).
  • SwarmFlow runtime controls, budget governance, and TUI commands – Added runtime control and budget-governance capabilities, together with a TUI command set for operational workflows (openJiuwen-ai/jiuwenswarm#2871).
  • SwarmFlow budget and nested workflow operators – Added budget operators and workflow nesting to support composable orchestration and execution constraints (openJiuwen-ai/jiuwenswarm#2182).

Symphony

  • Symphony atomic batch Skill graph updates – Added batch add, update, and delete interfaces that incrementally recalculate affected graph nodes and relations, publish immutable versions atomically, and support idempotency and concurrency validation (#715).

Evolution, Skills

  • AgentTemplate and Plugin loading – Added reusable expert templates and functional plugins with natural-language, form-based, and package-upload creation, runtime loading/unloading, legacy compatibility, path validation, conflict rollback, and composition (#268).
  • Structured agent execution traces – Added structured JSON collection of user requests, reasoning snippets, tool decisions, arguments, outputs, timestamps, and errors for session-level querying and diagnosis (#292).

Online RL

  • Online RL workflow improvements – Added trajectory CRUD APIs, LoRA/CKPT weight version management, LoRA metadata in trajectories, latest-hot-loaded-LoRA fallback, explicit training start/stop controls, and third-party Agent trajectory collection (#146).

🐛 Bug Fixes

Agent, Team & Runtime

  • Fixed interrupted-round streaming on the AgentOS branch (#1248).
  • Fixed hot-update tasks failing to evaluate completion correctly (#1228).
  • Fixed deferred tools failing approval interruption and resume (#1196).
  • Fixed closed subagents failing to resume after session-adapter recycling (#1185).
  • Fixed plugin tools disappearing after model hot updates (#1093).
  • Fixed team workspace synchronization and runtime adaptation issues (#1142, #1089).
  • Fixed team-mode _parse_level import error (#1021).
  • Improved browser-agent runtime management and synchronous tool binding (#1146).

Memory, Search & Skills

  • Optimized Symphony retrieval prompts (#1230).

  • Improved skill classification retrieval descriptions and Chinese/PPT skill conflict handling (#1129, #1054).

  • Fixed skill graph missing-model and absolute-path resolution messages (#1044, #1034).

  • Improved tool-search ordering, invocation recovery, and loop prevention (#1005, #998, #893, #766).

  • Added local short-read handling for Skill frontmatter (#843).

  • Fixed jiuwen_memory_provider compatibility with the latest jiuwen_memory package – Updated the provider integration to work with the current jiuwen-memory wheel and daily Swarm package (#1217).

  • Fixed mem2_search failing when top_k is omitted or None – Search now falls back to the default top-k value instead of calling int(None), restoring memory recall when the optional argument is absent (#731).

RSI, Artifacts & Models

  • Fixed program optimization evaluation, error classification, iteration loading, timeout handling, and event deadlocks (#1221, #1154).
  • Aligned artifact contracts, stage events, paper-model bridging, and publish progress (#1072, #1157).
  • Added the artifact optimization service (#1052).
  • Relocated RSI code and scaffolded artifact RSI components (#1031).
  • Refined no-proxy source fallback guidance (#1219).
  • Preserved paper evidence and task resources (#1207).
  • Fixed malformed LLM response field types being passed through (#995).
  • Improved model context-window discovery and default top_p handling (#891, #882).
  • Adapted Anthropic account integration to the upgraded version (#1097).

Security, Filesystem & Observability

  • Fixed security guards incorrectly denying operations in full-access mode (#1085).
  • Fixed predictable temporary paths exposing API keys and enabling path traversal (#364).
  • Fixed shared-host filesystem lock directory preventing other users from using filesystem tools (#282).
  • Fixed OTel trace breaks, missing component names, and None memory inputs (#1121).
  • Improved tokenizer-aware usage and KV-cache telemetry (#1063, #1009).
  • Fixed attachment placement affecting KV cache and model execution (#1011, #835).
  • Fixed Office document reads being blocked by ordinary text size limits (#973).
  • Improved image-modality probing for text-only models and cached rejection results (#498).
  • Improved browser subagent context engineering and webpage fetching (#809, #523).
  • Fixed branch synchronization and dev-stable integration issues (#503, #430, #878).
  • Hid internal tool names from team plans (#849).
  • Fixed SwarmFlow unexpectedly applying a token limit when budget was not requested (#1105).
  • Unified Agent-Core KV-cache lifecycle management for Ascend Affinity (#1102).

📚 Documentation

  • Added documentation for v0.1.18 capabilities – Documented SwarmFlow runtime and budget controls, team plan/reliability features, Permission Engine changes, Skill graph updates, AgentTemplate/Plugin loading, structured traces, and online RL APIs.

🔧 Experience & Tooling

  • Improved runtime operations and diagnostics – TUI controls, structured traces, consistent third-party-agent errors, and explicit RL training controls provide clearer operational feedback and safer day-to-day management.

🙏 Acknowledgments

Thank you to all contributors who submitted issues, design proposals, and validation feedback for v0.1.18. Your reports and testing continue to improve openJiuwen's reliability, governance, and extensibility.

Assets 5
Loading